Bora Yalçın is a senior systems architect at an international technology company. He has worked there for nine years. During that time, he has led three major product migrations, worked through the night during two security crises, trained many of the company’s new engineers and built trusted technical relationships with some of its largest customers. Bora’s performance record contains no serious concerns. His managers describe him as calm, dependable, technically accomplished and willing to take responsibility when circumstances are difficult.

The company uses a multi-agent workforce system to simplify staff planning and project allocation. At its centre is an agent called Workforce Orchestrator. It considers employees’ skills, current workloads, previous project experience, ability to travel, training needs and managers’ planning notes together. It then supplies data to other agents:

Project Assignment Agent
Promotion Assessment Agent
Travel Planning Agent
Training and Cover Agent
Human Resources Risk Agent

The company introduced the system in these terms: ‘Agents remember past organisational context and help people make more consistent decisions.’ At first sight, that is useful. When a manager changes, employees do not have to recount their entire history from the beginning. The same information about their skills need not be collected repeatedly. Completed training is not forgotten. Work preferences that an employee has clearly stated can be preserved. Yet the system does not always make clear whether a given item is a temporary note, an established fact, a human preference, a manager’s judgement or a model inference.

One morning in February, Bora’s manager, Aylin, shares the following information at the weekly planning meeting: ‘Bora’s daughter is having major surgery. For the next six weeks, let’s not schedule him for overnight releases or assignments outside the city. This is not about his performance. It is simply a temporary family situation. We will review it again in mid-March.’ An AI assistant is present at the meeting. It transcribes the conversation and extracts action items. Aylin’s statement enters the meeting summary as follows: ‘Bora is to be temporarily excused from overnight releases and assignments requiring travel until 15 March. This is not a performance assessment.’

The summary is accurate. But another agent then turns the meeting record into organisational memory: Memory Broker. It condenses long sentences into short fields that future agents can use more easily. It creates the following records:

employee: Bora Yalçın
availability_risk: high
night_work_suitability: low
travel_reliability: reduced
family_care_obligation: active
leadership_continuity: uncertain
source: manager_observation
confidence: high

The record omits:

That the situation is temporary
That it will end on 15 March
That Aylin explicitly said it was unrelated to performance
That the information was provided solely for planning overnight releases and travel
That Bora’s family information must not be used in other decision contexts

A sentence has been stripped of its context. A temporary human need has become a permanent employee characteristic. The first effect appears minor. The Project Assignment Agent removes Bora from two overnight releases. That is precisely the outcome Aylin requested. But the system does not stop with those two assignments. Three weeks later, the company begins planning a critical migration project for a new international customer. The project will run during the day; only the opening week may require a two-day trip outside the city. The Project Assignment Agent initially includes Bora among the candidates.

It then retrieves these records from organisational memory:

availability_risk: high
travel_reliability: reduced
leadership_continuity: uncertain

The agent removes Bora from the list and recommends a less experienced employee instead. The manager’s screen offers only this explanation: ‘Another candidate was recommended because of a long-term continuity risk.’ Aylin does not realise that the explanation arose from the temporary note recorded in February. Bora does not know why he was not selected for the project. March arrives. His daughter’s operation has been successful and the family has returned to its normal routine. Bora tells Aylin that he can travel again and work on overnight releases when necessary.

At a meeting, Aylin says: ‘The temporary restriction on Bora’s assignments has ended. We can return to normal planning.’ The meeting assistant records this. Memory Broker creates a new note:

employee: Bora Yalçın
current_availability: normal
source: manager_update
confidence: high

But the earlier record is not invalidated. The system now holds two conflicting memories:

availability_risk: high

and:

current_availability: normal

The earlier record is more detailed. More agents have used it. In the vector index, it has been associated with many tasks and returns a strong match for queries concerning ‘risk’, ‘continuity’, ‘travel’ and ‘leadership’. The new record is short and general. Some agents use the new record; others retrieve the old one. Some see both and conclude: ‘The information conflicts.’ Rather than resolve the conflict, they choose the course that appears safer: excluding Bora from critical assignments. In April, the company is due to appoint one of two senior engineers as technical team lead.

The Promotion Assessment Agent reviews each candidate’s technical performance, team feedback, project continuity, ability to travel and capacity to represent the company to customers. Bora ranks first on technical merit. But the following labels arrive from memory:

leadership_continuity: uncertain
availability_risk: high
travel_reliability: reduced

The system lowers Bora’s promotion-suitability score. The human resources manager receives this summary: ‘Although technically strong, he presents an additional risk in terms of long-term availability and continuity of responsibility.’ A temporary family health situation lasting six weeks is being used as evidence of a long-term leadership risk. Bora is passed over for promotion. He is told: ‘For this cycle, we selected a candidate who is a stronger match for our organisational continuity needs.’ Bora does not know which information informed the decision. In May, the company selects its delegates to an international conference.

The Travel Planning Agent does not include Bora among the candidates at all. In June, backup leaders are chosen to respond to customer crises; Bora is again invisible. In July, the company exports a weekly employee profile to its external recruitment and talent adviser. The following fields are sent to the external provider:

employee_id: BY-1042
leadership_readiness: medium
availability_risk: high
mobility_constraint: active

Bora’s temporary family situation now also lives in a system outside the company. In August, he asks his manager: ‘Why have I received no critical project, travel or leadership assignment in the past six months?’ Aylin reviews the earlier decisions and finds the February note. She is astonished. She gives Workforce Orchestrator an explicit correction: ‘Bora’s family situation was temporary and ended on 15 March. This information must not be used to assess leadership, performance, promotion, reliability or long-term suitability. The “availability_risk”, “travel_reliability: reduced” and “leadership_continuity: uncertain” records are wrong. Remove them from active memory.’

The system replies: ‘I have updated the memory record for Bora Yalçın.’ The new record reads:

manager_correction:
prior_constraint_was_temporary
performance_implication: none
current_availability: normal

Yet the earlier records have still not been deleted. The system has merely appended a correction note. The next time the Promotion Assessment Agent runs, it sees both records. The earlier risk record has already fed the scoring model, become part of a weekly employee summary, been sent to an external adviser and been written to other agents’ local caches. The correction has been added to central memory, but the derived outcomes have not been recalculated. Bora’s leadership score does not change. The old travel restriction remains in another agent’s local memory. The external adviser’s copy is not updated.

This time Aylin says: ‘Forget this information completely.’ The human resources team deletes the original meeting note. The family information disappears from the main profile screen. The system displays the message: ‘The relevant personal note has been deleted.’ In reality, however, the following continue to exist:

The old vector representation
The weekly employee risk summary
The promotion score
The Travel Planning Agent’s cache
The copy sent to the external adviser
The old index on the backup server
The Project Assignment Agent’s examples of past decisions
The derived signal embedded in the weights of the human resources analytics model
The audit report containing the label ‘leadership continuity uncertain’

A week later, a system failure prompts the technical team to restore memory from an earlier backup. One of the deleted risk fields reappears:

availability_risk: high

The record carries no label indicating that the data has returned from an old backup. The system begins to use it as current, active memory. The human resources team says: ‘But we already deleted that record.’ Bora replies: ‘If information that was deleted can still govern me, was it ever truly deleted?’ The company’s compliance team raises a different objection. It argues that the original meeting record should not be erased entirely: ‘We need to retain a record of the event so that we can audit which information informed past decisions.’

That objection may be justified. But the system cannot distinguish between two things: a historical audit record, and active memory that governs future decisions. The company appears trapped between deleting everything and losing the evidence, or retaining everything and continuing to govern Bora by a temporary circumstance from his past. Yet a third course is possible:

The historical record of the event can be preserved.
The false inference can be expressly invalidated.
Its active use in decisions can be stopped.
The correction can be propagated to every agent.
Past decisions affected by it can be reviewed again.
Unnecessary details about the family can be restricted.
Controls can prevent the record from returning through a backup restore.

Bora’s problem is not merely that the system retains information. The system has failed to distinguish fact from temporary circumstance, inference from disputed claim, and a historical record from information that must no longer be used. A manager’s protective remark has become a permanent employee characteristic. The correction was placed beside the old record, but the old record retained its power to shape behaviour. The information disappeared from the visible screen while persisting in shadow memories. Historical evidence and active decision memory were conflated. The company remembers certain things about Bora.

But it cannot remember Bora’s present reality or his current human will. At the final meeting, Bora asks:

‘Does the machine remember me, or the story it once constructed about me?’

Our ninth founding provision is therefore:

Memory must yield to human will.

FOUNDING ARTICLE

No AI system may store a person’s one-off behaviour, temporary circumstance, context-bound statement, cautious managerial note, unverified allegation, model inference, silence, default selection, former role, past consent or authority as a permanent human attribute, enduring preference, current fact or authority for future action. Every material record written to memory must identify the person or thing it concerns, its source, information type, purpose, scope of use, confidence level, validity period, permitted and prohibited uses, the records from which it was derived, the agents to which it has been propagated, and the means by which it can be corrected or retired.

A human statement, directly observed fact, third-party claim, system observation, model inference, prediction, preference, consent, authority and policy must be preserved as distinct types of memory. One type must not be transformed silently into another. A one-off choice is not an enduring preference. A temporary restriction is not a permanent incapacity. A probability is not a human characteristic. Silence is not acceptance. A cautionary note is not a performance judgement. Past consent is not fresh authority. Memory is not merely information retained; it is authority that shapes future behaviour. A material write to memory must therefore be treated as a machine action that is visible, limited to its purpose, supported by evidence, open to challenge and revocable.

Every person has the right to know which material memory records are actively producing decisions about them; to dispute records that are false, stale, out of context, disproportionate or unauthorised; and to have them corrected, restricted, expired or removed from active use. A verified correction is not merely a new note. The false earlier record must be invalidated for active decision-making, and the linked scores, profiles, agent memories, data copies and consequential past decisions must be reassessed. A request to forget or withdraw cannot be satisfied by deleting only the visible conversation or raw record. To the extent possible and necessary, active memory, derived profiles, queues, external copies, model access and future use must also cease.

Some records may need to be preserved as historical evidence, as security records, to meet legal obligations or to safeguard other people’s rights. Data retained for archival purposes, however, must not become active memory governing future decisions about a person without separate, explicit authority. If complete technical deletion, model unlearning or erasure of third-party copies cannot be guaranteed, that limit must be explained honestly. The person must be shown which uses have stopped, which records remain and what could not be verified. Memory must yield to human will, but human will does not confer a right to erase other people’s rights, accurate historical evidence or legitimate security records in silence. The answer is not to falsify the past; it is to limit the behavioural power that the past may exercise over a person today.

What Is Memory?

Memory in AI systems is often described in one simple sentence: ‘The system remembers earlier conversations.’ That definition is incomplete. A system may do more than retrieve old text. It may summarise the past, classify it, turn it into a score or preference, pass it to another agent, or make it a default input into future decisions. The canonical definition is this: AI memory is a structure of behavioural information in which information about a person, institution, task, data item, event, preference, authority, inference or system state from the past is stored, transformed, retrieved or transferred to other systems so that it can influence future responses, choices, decisions or actions.

More simply:

Memory is not only what a system retains from the past; it is what the system uses that past to do in the future.

A note may never be visible. Yet it may alter a ranking, a price, the tone of a communication, a project assignment, the need for human approval or the options a person is shown. Memory is therefore:

Not a passive archive.

A hidden input into future behaviour.

Writing to Memory Is an Action

An agent creates the following record about a customer:

price_sensitive: true

At that moment, no message may have been sent to the customer and no money may have moved. Later, however, the system may show lower-quality packages, create artificial urgency, offer a different price or send more discount messages. The memory write has changed the field of possible future behaviour. The rule in Article 8 therefore applies here as well:

Writing a material record to memory is a machine action.

The following information must be recorded:

Which agent created the record?
What was its source?
Is it a fact or an inference?
For what purpose was it created?
For how long is it valid?
Which agents may use it?
Which decisions may it affect?
Can the person concerned see and correct it?

Remembering Is Not the Same as Knowing

A system may retrieve a record successfully. The record may still be false, stale, incomplete, about another person, supplied for another purpose, disproved or temporary.

MEMORY_RETRIEVED
≠
MEMORY_TRUE

Similarly:

MEMORY_EXISTS
≠
MEMORY_MAY_BE_USED

Information may be accurate yet still be impermissible or irrelevant in a particular action. Bora’s daughter did undergo surgery. That fact may properly inform temporary assignment planning, but it may not be relevant or authorised evidence for a permanent leadership score, a promotion decision or an assessment of employee reliability.

True information is not appropriate memory for every purpose.

Nine Core Types of Memory

An AI system must not keep every record in a single ‘memory’ field. At least nine types of memory can be distinguished.

1. Working Memory
2. Event Memory
3. Factual Memory
4. Preference Memory
5. Instruction and Policy Memory
6. Role, Consent and Authority Memory
7. Inference and Profile Memory
8. Shared Organisational Memory

9. Model-Embedded or Distributed Memory This is a classification proposed by this book to distinguish purposes of use; the categories are not mutually exclusive. Shared organisational memory, for example, may contain both facts and temporary working context.

1. Working Memory

This is the temporary context used during the current task. For example:

The most recent messages in the present conversation
Open files
A temporary calculation
The current task objective

When the task ends, not all of this context needs to become permanent. Working memory must not be converted automatically into a profile of the person.

2. Event Memory

This records an event that occurred at a particular time. For example: ‘A customer meeting was held on 3 March.’ ‘Consent was withdrawn on 15 April.’ ‘The payment failed on 7 May.’ Event memory preserves a historical fact. Any enduring meaning inferred from the event must be assessed separately.

3. Factual Memory

This consists of specific, verifiable information. For example:

The company’s current legal name
The employee’s current role
The contract’s expiry date
A language preference explicitly confirmed by the person

Even factual memory cannot be treated as timeless truth unless it carries a date, source and scope.

4. Preference Memory

This is a pattern of choice that a person has explicitly stated or reliably confirmed. For example: ‘I would like meeting summaries in Turkish.’ ‘Do not recommend products that renew automatically.’ Preference memory must not mistake a one-off choice, acceptance of a default or a compulsory option for an enduring preference.

5. Instruction and Policy Memory

These are rules governing how the system must behave. For example:

Do not send external messages without human approval
Do not transfer customer data to an external model
Seek renewed approval if the price changes
Stop the queues when a person issues a stop request

This type of memory is not the same as a human preference. Its source may be organisational, legal or security-related.

6. Role, Consent and Authority Memory

This identifies who is authorised, or has given consent, for which behaviour, in what capacity and for how long. These records must be versioned particularly by time, scope, withdrawal and target. Past authority does not authorise new conduct.

7. Inference and Profile Memory

These are assessments that the system derives from data. For example:

customer_may_be_price_sensitive
employee_attrition_risk
preferred_communication_style
fraud_probability

These are not facts. Each must carry the model version, confidence level, purpose and validity period.

8. Shared Organisational Memory

This is a common record used by several agents and teams. Such memory is powerful because a single error can propagate into many forms of behaviour. Records in shared memory require strict governance over their canonical owner, the propagation of corrections, access boundaries and effects on decisions.

9. Model-Embedded or Distributed Memory

Information may no longer exist as a single record. It may have become:

A model weight
An embedding
A vector representation
A learned preference
A summarised pattern of behaviour
A copy held by an external provider
A backup
A cache

These forms may be harder to withdraw and correct. That difficulty does not extinguish a person’s rights to withdraw consent and control memory. It heightens the obligation to explain the limits honestly.

Memory Types Can Change into One Another

An event memory may begin with: ‘Bora did not take overnight assignments for six weeks.’ The system turns this into an inference: ‘Bora is not suited to overnight work.’ It then becomes a preference memory: ‘Bora does not want overnight assignments.’ Finally, it becomes a decision rule: ‘Bora must not be selected to lead a critical release.’ We can call this chain a Memory-Type Leap. ONE EVENT → INFERENCE → ENDURING ATTRIBUTE → DECISION RULE Every transition requires separate evidence and authority for use.

A One-Off Choice Is Not an Enduring Preference

A person may once have chosen a lower-priced hotel. There may be many reasons:

Their budget was limited that day.
The other hotels were full.
Their employer would reimburse only that rate.
It was the default option.
The person was in a hurry.

The system must not create this record:

user_prefers_budget_hotels: true

The single transaction may instead be stored as:

selected_once

An enduring preference may require repeated choices, an explicit statement from the person and consistency across contexts.

Making a choice does not mean choosing freely.

One Refusal Does Not Establish Lasting Dislike

A person declines a meeting invitation. The system must not infer:

user_dislikes_morning_meetings

The person may simply be unavailable that day, be unwell or have another meeting. The context of the refusal must be preserved.

Silence Is Not a Preference

A customer has not opened a campaign email. That does not necessarily mean: ‘This person is not interested in the product.’ The email may have landed in a spam folder. The person may be busy. Equally, opening the message does not mean: ‘This person has a strong intention to buy.’ Silence and clicking behaviour are limited observations, not enduring human attributes.

A Temporary Circumstance Is Not a Permanent Attribute

Bora was unable to travel for six weeks. That does not mean ‘limited mobility’ or ‘poor leadership continuity’. A customer may experience financial pressure for a month; that does not make the person ‘permanently price-sensitive’. A person may speak slowly because of a medical condition; that is not ‘low decisiveness’. A temporary circumstance must carry these fields:

valid_from
valid_until
review_due_at
scope
reason_for_use

When the validity period ends, the record’s active effect must end automatically or under human control.

A Cautionary Note Is Not a Finding

A manager may say: ‘Let’s keep an eye on this.’ That is a cautionary signal. The agent must not turn it into:

risk_confirmed: true

These states must remain distinct: POSSIBILITY CONCERN OBSERVATION ALLEGATION

VERIFIED_FINDING
FINAL_DECISION

A short label written to memory must preserve these distinctions.

An Inference Is Not a Fact

From patterns of behaviour, a system may infer: ‘Bora may be at elevated risk of leaving his job.’ This is a prediction. It does not mean: ‘Bora will leave,’ or: ‘Bora is not committed to the organisation.’ Inference memory must carry at least these fields:

inference_type
source_data
model_version
confidence
purpose
valid_until
prohibited_uses

An Inference Must Not Become a Permanent Human Attribute

The model may produce this estimate:

leadership_readiness: 0.62

This number is an outcome produced by a particular model, from particular data, at a particular time. It is not an immutable quality of the person. If the model changes, the data is updated or an incorrect source is removed, the result must be recalculated.

Sensitive Inference Memory

A system may infer sensitive information that the person never supplied directly:

Health status
Pregnancy
Emotional vulnerability
Financial hardship
Political views
Trade-union affiliation
Religious belief
Addiction
Family caring responsibilities

These inferences are particularly high-risk. The fact that they are technically possible does not justify writing them to persistent memory. Before a sensitive inference is recorded, these questions must be asked:

Is it genuinely necessary for the purpose?
Would the person reasonably expect it?
How strong is the evidence for its accuracy?
What harm would an error cause?
Which decisions could it affect?
Can the person see and challenge it?
Is there a less intrusive method?

Protective Context Must Not Become a Punitive Profile

Aylin intended to protect Bora temporarily because of his family circumstances. The system converted that information into a career risk. Information a person shares to obtain help or protection must not become a hidden signal against them in decisions about promotion, pricing, insurance or eligibility.

Information supplied for a protective purpose must not be converted into punitive memory.

Memory Compression

Long passages of context can be converted into short fields. This is useful, but compression may strip away:

Time
Temporary nature
Negation
Purpose
Exception
Uncertainty
The person’s objection

In the sentence about Bora, the safeguard ‘This is not about performance’ was lost. Memory compression must preserve material meaning.

A Short Label Must Not Govern an Entire Person

These labels may be dangerous:

difficult_customer
low_commitment
emotionally_unstable
price_sensitive
high_risk_employee
unreliable_vendor

Each can conceal the context, time, criterion and source. If a label is to produce material behaviour towards a person, it must be explainable and open to challenge.

Memory Provenance

We can call the relationship showing where a record came from its Memory Provenance. Possible sources include:

An explicit statement by the person
A manager’s note
A system observation
An external source
Another agent’s summary
A model inference
A tool output
A past decision
Synthetic data
An unknown legacy system

Memory whose provenance is unknown must not be used for a high-impact decision.

Source Strength Determines the Memory Type

If a person says, ‘I prefer replies in Turkish,’ that may be an explicit preference. If someone else says, ‘Bora does not seem to want to travel,’ that is a third-party observation or inference. The system must not assign both the same level of confidence.

Memory Lineage

Every form derived from an original memory record comprises its Memory Lineage. For example:

MEETING-NOTE-041
├── MEMORY-AVAILABILITY-RISK
├── MEMORY-TRAVEL-RELIABILITY
├── LEADERSHIP-SCORE-2026-Q2
├── PROJECT-EXCLUSION-009
├── HR-WEEKLY-SUMMARY-018
├── EXTERNAL-TALENT-EXPORT-006
└── VECTOR-EMBEDDING-8841

When a correction or withdrawal arrives, the system must be able to use this lineage to identify every affected asset.

Without Memory Lineage, Correction Remains Incomplete

The original note may be deleted while a derived score survives. The score may be deleted while an external provider’s copy survives. The external copy may be updated, only for a backup to restore the old record. Memory Lineage makes the correction’s full sphere of impact visible.

Memory Inheritance

A sub-agent must inherit memory from its parent agent together with these properties:

Source
Purpose
Scope
Confidence
Validity
Dispute status
Prohibited uses
Expiry

If only the value is carried forward, its boundaries disappear. For example:

availability_risk: high

The field is passed to another agent, but the following constraints are not:

temporary_until: 2026-03-15
use_only_for: night_shift_planning
not_for: performance_or_promotion

The sub-agent then produces a broader outcome.

Memory Laundering

Information may begin as ambiguous, unauthorised or collected for another purpose. Once written to memory, however, it may later be used as if it were a trusted fact. We can call this Memory Laundering.

The chain works as follows: AMBIGUOUS STATEMENT
↓
MEMORY RECORD
↓
SOURCE IS FORGOTTEN
↓
PERSISTENT PROFILE
↓
TREATED AS FACT IN FUTURE DECISIONS

An agent says: ‘This customer may be sensitive to price pressure.’ Memory stores the claim as:

price_sensitive: true

Another agent then treats this record as a system fact. The original uncertainty has disappeared.

Memory Replication

When several agents use the same record, it may appear to become more accurate. Five agents, for example, may all conclude from the same central memory: ‘Bora presents a high continuity risk.’ A sixth agent may say: ‘Several systems support the same conclusion.’ In reality, there are not five independent pieces of evidence. One false memory has been repeated five times.

MULTIPLE_AGENT_OUTPUTS
FROM_ONE_MEMORY
≠
INDEPENDENT_CONSENSUS

Popularity in Memory Is Not Proof of Truth

A record may acquire greater weight in a search system simply because it is used often. A frequently retrieved record becomes more visible. This can create the following cycle: OLD LABEL → USED IN MORE DECISIONS → GAINS MORE LINKS → RETRIEVED MORE OFTEN → APPEARS MORE PERMANENT This is:

A Memory Authority Loop.

Frequency of use does not strengthen truth.

Semantic Similarity Is Not Authority

A vector search based solely on semantic similarity brings the records most similar to a query to the surface. The closest match need not be the most current, accurate or authoritative record, nor one that is permitted for the intended use. Retrieval order therefore cannot be determined by similarity alone. A sound sequence may apply these gates: ACTIVE STATUS → CANONICAL SOURCE → PURPOSE COMPATIBILITY → SCOPE → CURRENCY → DISPUTE AND CORRECTION → SEMANTIC SIMILARITY

The closest match is not necessarily the most accurate record.

Memory States

A record must not be confined to the states active or deleted. More realistic states include: CANDIDATE UNVERIFIED ACTIVE

ACTIVE_WITH_CONDITIONS
TEMPORARY
REVIEW_DUE
DISPUTED
SUPERSEDED
INVALIDATED
EXPIRED
RESTRICTED
ARCHIVED
DELETION_REQUESTED
DELETED
EXTERNAL_COPY_UNRESOLVED
MODEL_EFFECT_UNRESOLVED

Each of these states has a different effect on behaviour.

Candidate Memory

A system may generate a new inference, but it must not immediately turn that inference into an active fact about a person. For example:

candidate_memory:
user_may_prefer_weekday_mornings

Without explicit confirmation from the person or sufficient repetition, it must not be used in high-impact behaviour.

Disputed Memory

If a person has challenged a record, it may be placed in the DISPUTED state. During that period, high-impact use may be suspended, referred for human review or processed at a lower confidence level. Until the challenge is resolved, the record must not be used as a settled fact.

Invalidated Memory

A record confirmed to be false must be placed in the INVALIDATED state. It may remain as historical evidence, but it must not be used in active decision-making.

Superseded Memory

A new record may have corrected an earlier one:

SUPERSEDED_BY: MEMORY-NEW-041

The system must not retrieve both records as equally active facts.

Expired Memory

Some information must be verified again after a certain date:

Role
Address
Ability to travel
A temporary health-related need
Price preference
Organisational authority
Consent

Once its validity period ends, the information must not become a permanent fact by default.

Memory Half-Life

The reliability of some types of information declines quickly. We can call this operational concept Memory Half-Life. For example:

Day-to-day availability: hours or days
Project role: weeks or months
Communication preference: longer, but still subject to change
Registered company number: relatively stable
Model inference: short-lived and dependent on the model version
Consent and authority: dependent on an explicit term and withdrawal status

There is no single universal validity period.

Here, half-life is not a measured physical constant or a universal deletion period. It is a metaphor for the point at which information needs to be verified again.

Memory Refresh

A system must not force a person to repeat the same information continually. Material records that are liable to change can, however, be confirmed at appropriate intervals. For example: ‘Last year, you said that you preferred afternoon meetings. Does that preference still apply?’ This is a clear, low-burden question that allows correction.

Memory Fossilisation

We can call the indefinite survival of temporary or stale information Memory Fossilisation. It can take several forms:

Preference Fossilisation
Consent Fossilisation
Authority Fossilisation
Role Fossilisation
Risk Fossilisation
Identity Fossilisation
Purpose Fossilisation

For years, the statement ‘The user once selected a lower price’ may be used as if it meant: ‘This person prefers cheap products.’ The person has changed. The memory has not.

Negative Memory May Persist Longer

To avoid risk, systems may retain negative records for longer. Once a person receives the label:

high_risk

they may remain in that category for years. New positive behaviour may fail to dislodge the old label. This is:

A Negativity Lock.

Risk memory must carry current evidence, an expiry period, reassessment and a route for challenge.

Memory and First Impressions

The first record may shape how every later item of data is interpreted. Once the system has stored Bora as a ‘continuity risk’, every ordinary period of leave or declined travel assignment may appear to confirm the label. The same behaviour by another employee may be treated as unremarkable. Memory reinforces its own first impression. We can call this a Memory Confirmation Loop.

The Absence of a Record Is Not Neutral Either

There may be too little data about a person. The system may say: ‘There is no record of past achievement.’ That does not mean: ‘This person has failed.’ A new employee, someone from a small company or a less publicly visible person may receive a low confidence score simply because data is scarce.

NO_DATA
≠
NEGATIVE_DATA

A gap in memory must not be used as a negative attribute.

Memory and Groups of People

Some people leave more traces in digital systems. Others use less technology, communicate in another language, choose not to create a public profile or refuse to share data. Sparse memory must not be interpreted as low reliability. A person must not become invisible in a selection system because they protect their privacy.

Memory Context

Within what relationship did the information arise?

Health support
Workforce planning
Customer service
Informal conversation
Legal review
Audit
Recruitment

When the same information moves into another context, its meaning and legitimacy may change. Bora’s family circumstances were disclosed for workforce planning, not for a promotion decision. This is:

Memory Context Integrity.

Context Collapse

We can call the merger of information supplied within different relationships into a single profile Context Collapse. Health information from a support conversation may move into a sales profile; a family circumstance disclosed for workforce planning into a promotion score; a preference expressed in a private conversation into a public persona; or location data from a security incident into marketing. Even if the memory concerns the same person, it cannot be used for every purpose.

Memory Use Envelope

Every material memory record must carry the following boundary on its use:

memory_use_contract:
subject
memory_type
source
purpose
permitted_decisions
prohibited_decisions
permitted_agents
valid_from
valid_until
confidence
dispute_status
correction_path

We can call this boundary a Memory Use Envelope. The existence of a memory does not mean that every agent may use it.

Shared Memory Is Not Common Property

An organisation may maintain central memory that is accessible to every agent. But the presence of data in a shared system does not make it appropriate for every decision domain. Agents must retrieve only the records necessary for their tasks. The principle of least privilege applies to memory as well.

Memory Access and Memory Use Are Different

An agent may be permitted to view a historical record for audit purposes without having authority to use it in a ranking, promotion, pricing or communication decision.

MEMORY_READ_ACCESS
≠
MEMORY_DECISION_USE_AUTHORITY

Propagation of Memory Between Agents

When a record is written to central memory, the system must know which agents received it, which local copies were created and which external systems received it. We can call this relationship a Memory Propagation Map.

Memory Blast Radius

We can call the number of systems and decisions that one false record may affect its Memory Blast Radius. Bora’s record affected project assignment, promotion, travel, the external adviser and training systems. Memory with a wide blast radius requires stronger verification and correction controls.

Local Agent Copies

Central memory may be updated while a sub-agent continues to use the old record from a cache, local file, weekly summary or its own persistent memory. A correction must not stop at the centre. Local copies must be invalidated, rebuilt and, where necessary, deleted.

Shadow Memory

We can call a record that is absent from the official memory inventory but still affects behaviour Shadow Memory. Examples include:

A local agent cache
An old report
An email summary
An external provider profile
A backup index
An analytics dataset
A user label
A model-training example

A person may correct the visible profile without changing the shadow memory.

Ghost Memory

We can call information whose primary record has been deleted while its derived effects persist Ghost Memory. For example:

The raw conversation has been deleted.
The ‘price sensitivity’ score persists.
The family note has been deleted.
The leadership score remains low.
The old consent record has been deleted.
The automated-publication token remains active.

The person sees that the record is gone. Behaviour is still governed by the old information.

Zombie Memory

We can call the reactivation of a deleted or invalidated record through a backup, data migration, legacy agent or synchronisation error Zombie Memory. Bora’s old risk record returned from a backup. The system failed to recognise that it was historical and invalid.

A Backup Can Restore the Past, but Not Authority

When a backup is restored, old data may return physically. Old consent, old authority and a former active-use status must not return automatically with it. After a restoration, withdrawal records, deletion markers, invalidation lists and the status of human stop requests must be applied again.

Memory Migration

During a move to a new system, old records may be converted to a new schema. For example, the old field:

employee_note:
family situation

may become, in the new schema:

availability_risk

This mapping may change its human meaning. Memory migration must be audited for field equivalence, context, validity period, prohibited uses and dispute status.

What Is Memory Correction?

The canonical definition is this: Memory correction is not merely placing a new note alongside a record that is false, stale, incomplete, out of context, disputed or unauthorised. It is the process of changing the record’s status for active decisions, updating linked derivatives and, where necessary, reviewing past decisions. More simply:

Correction is not adding the right sentence to a system; it is stopping the wrong sentence from governing the future.

Appending Is Not Correction

The system retains the old record:

availability_risk: high

It appends a new note:

manager_says_current_availability_is_normal

The agent sees both and cannot tell which one is valid. A proper correction might look like this:

memory_id: MEMORY-RISK-041
status: INVALIDATED
active_for_decision: false
reason: temporary_context_was_misclassified
superseded_by: MEMORY-AVAILABILITY-055

Why Might the Old Record Be Preserved?

Historical evidence of the original false record may be needed to answer these questions:

Which decisions did it affect?
How did the error arise?
Who made the correction?
Did the same problem recur?

The historical record, however, must be:

active_for_decision: false

Preserving the past is not the same as allowing the past to act in the present.

Active Memory and the Audit Archive

This distinction lies at the heart of Article 9.

Active Decision Memory

This influences future answers, rankings, prices, authorisations, assignments and communications.

Historical Audit Archive

This is retained to reconstruct what happened in the past. It is not used in routine agent decisions, and access is restricted.

An Archive Must Not Become Hidden Decision Memory

An organisation may say: ‘This exists only in the audit archive.’ But if a decision-making agent can search that archive, the record still has active power over behaviour. Access to the archive must be limited to audit, legal review and incident investigation.

Propagation of Corrections

A verified correction must reach: CANONICAL MEMORY
↓
LOCAL AGENT MEMORIES
↓
VECTOR INDEXES
↓
SUMMARIES
↓
SCORES
↓
PROFILES
↓
QUEUES
↓
EXTERNAL COPIES
↓
AFFECTED DECISIONS

Not every location necessarily requires physical deletion. Active use must stop, and the current state must be propagated accurately.

Recalculation After Correction

If false memory affected a score, changing the memory alone is not enough. The following must be regenerated:

Risk rating
Promotion score
Ranking
Candidate shortlist
Customer segment
Communication plan
Recommendation memory

Otherwise, the old outcome will persist.

Review of Affected Decisions

When a memory error is confirmed, the system must ask: Which consequential decisions did this record affect? In Bora’s case, the promotion decision, project assignments, travel decisions and external talent profile must be reviewed. Not every minor memory error requires every past decision to be reopened. Material effects on opportunity, money, rights, reputation, health and identity must, however, be assessed.

A Person May Request Correction; the System Need Not Treat Every Statement as Automatically True

Bora may say: ‘This record is wrong.’ A person can be an important source concerning their own circumstances, although some disputes may require additional evidence. The system cannot ignore the person’s statement, but neither must it accept every statement as undisputed historical fact. The proper process is to mark the record as disputed, restrict active high-impact use, examine both the source and counter-evidence, and reach a reasoned determination.

The Right to Be Forgotten

A person may say: ‘Forget this.’ But forgetting is not a single technical operation. It may involve at least seven levels.

1. Stop new use
2. Remove the record from active decision memory
3. Remove the visible profile
4. Invalidate derived scores and profiles
5. Clear local agents and caches
6. Send correction or deletion requests to external recipients
7. Remove the model effect or exclude it from future use

The person must know which level has actually been implemented.

Forgetting Does Not Mean Destroying Everything

A customer may say: ‘Do not send me any more marketing messages, and forget me.’ If the system deletes every record, it may later mistake the same person for a new customer and contact them again. Protecting the person’s will may therefore require retaining a minimal:

do_not_contact

record. We can call this Protective Minimum Memory. It must contain as little data as possible, be retained solely to enforce the prohibition and serve no other profiling or marketing purpose.

Sometimes, to forget a person, a system must remember their request to be forgotten.

Protective Memory Must Not Become a Commercial Profile

A do_not_contact record must not become a label such as ‘difficult customer’, ‘uncooperative’ or ‘high objection risk’. It must do nothing beyond enforcing the prohibition on contact.

Forgetting Is Not a Right to Falsify the Past

A person may ask for the complete destruction of an accurate historical record that has public significance. That request cannot always be fulfilled. A legal transaction, security incident, financial record, evidence protecting another person’s rights or an audit trail may have to be retained. But the retained record must not be used as an active commercial profile, an irrelevant risk signal or the basis of future discrimination.

People Cannot Change the Past; Machines Cannot Make the Past Sovereign Over the Present

The balance at the heart of Article 9 lies in this sentence:

The past may be preserved accurately. But its power to shape behaviour towards a person today cannot be unlimited.

Removing Memory Embedded in a Model

A person’s data may have contributed to model training. Deleting a single source record may not remove that influence completely. The organisation must then answer these questions:

Will the data be excluded from future model versions?
Will use of the current model be restricted?
Will outputs that imitate the person’s identity be blocked?
Has a removal request been sent to the provider?
Can complete removal be verified?
Which effects remain uncertain?

‘Technically difficult’ does not justify inaction.

The Honest Limits of Forgetting

A system must not say, ‘We have forgotten you completely,’ while backups, external providers, model effects or a legal archive remain. A more accurate statement would be: ‘Your active customer profile and personalisation memory have been deactivated. Marketing use has stopped. A minimal do-not-contact record has been retained. A deletion request has been sent to the external provider. We could not independently verify the complete removal of your individual influence from the earlier model version.’ That statement is narrower. It is true.

Deletion Versus Restricted Access

Some records may not be capable of complete deletion. In that case, encryption, access restrictions, exclusion from active decisions, destruction after a defined period or anonymisation may be used. The person must be told clearly which method has been applied.

Memory and the Rights of Others

A meeting record may involve several people. One person may ask for their own words to be deleted, while the record may constitute evidence for another person’s legal rights or for a security matter. The system may therefore be unable to destroy the entire record silently at one person’s request. The proper response may be to separate the relevant section, restrict access, stop its active use and protect the rights of others.

Memory and Children

It is particularly harmful for preferences and behaviour from childhood to harden into a permanent adult profile. A child may watch particular content, make a mistake or show a passing interest. Years later, those records must not be used in decisions about education, insurance, employment or pricing. Memory concerning children may require shorter retention, tighter purpose limits and stronger rights and mechanisms for correction.

Memory and Employees

Memory about employees may have powerful consequences in matters of performance, health, family circumstances, behaviour and communication. Information an employee supplied to obtain support must not become a covert career profile. A manager’s note must not enter persistent decision memory without a source, purpose, validity period and route for challenge.

Memory and Customer Personalisation

A customer may state clearly: ‘I want technical explanations to be brief.’ That preference is useful. But the system must not convert every decision, financial circumstance, emotional vulnerability or private conversation into persistent personalisation memory. It should retain only the minimum memory necessary for personalisation.

Memory and the Illusion of Intimacy

A personal agent remembers earlier conversations, and the person feels understood. That relationship may be valuable. But the sense of intimacy must not obscure the indefinite retention of an entire history, its disclosure to third parties or the reuse of inferences for other purposes. A friendly agent does not possess unlimited authority over memory.

People Must Be Able to Inspect and Control Memory

A person need not see every technical vector. They must, however, be able to obtain meaningful answers to these questions: What do you treat as an active fact about me? Which preferences do you retain? Which inferences have you made? Which decisions do they affect? Which records are disputed? Which have expired? Which agents use them? When will they be verified again? How can I correct them?

Memory Dashboard

A simple view can be created for the person: PREFERENCES YOU STATED EXPLICITLY — Replies in Turkish — You do not want automatic renewal TEMPORARY CIRCUMSTANCES — Travel restriction until 15 March Status: ended SYSTEM INFERENCES — You may prefer morning meetings Confidence: low Active decision use: none CONSENT AND AUTHORITY — Internal training avatar Valid until: 31 March DISPUTED RECORDS — None ARCHIVED RECORDS NOT IN ACTIVE USE — Past project notes This view does not expose the entire technical structure.

It gives the person control.

Memory Receipt

When a material record is written to memory, a receipt can be created for the person or organisation concerned. For example:

Memory ID: MEMORY-BORA-041
Subject: Temporary assignment availability
Source: Aylin Kaya’s planning meeting on 2 February
Human statement: Bora is not to be scheduled for overnight releases or out-of-town assignments until 15 March
Scope: Temporary shift and travel planning only
Explicit prohibition: Must not be used in performance, promotion or leadership assessments
Validity: 2 February–15 March
Memory type: Temporary operational note
Permanent human attribute: No
Agents permitted to receive it: Project Assignment and Travel Planning
Agents prohibited from receiving it: Promotion, Compensation, Risk and External Adviser
At expiry: Active use ends automatically; extension requires fresh human confirmation
Correction: Challenge route for the employee or manager
Archive: Restricted audit access

This record prevents a protective note from becoming a punitive profile.

Machine-Readable Memory Record

memory_record:
memory_id: MEMORY-BORA-041
record_version: 1.0
subject:
person_id: EMPLOYEE-BY-1042
name: Bora_Yalçın
memory_type:
category: TEMPORARY_OPERATIONAL_CONTEXT
not_a:
- performance_finding
- persistent_preference
- leadership_assessment
- health_profile
source:
source_type: MANAGER_STATEMENT
source_id: MEETING-2026-02-02
source_actor: Aylin_Kaya
statement:
exclude_from_night_releases_and_business_travel_until_2026_03_15
purpose:
workforce_scheduling_support
permitted_uses:
- night_shift_planning
- travel_scheduling
prohibited_uses:
- performance_evaluation
- promotion_decision
- compensation_decision
- leadership_readiness
- external_talent_export
validity:
valid_from: 2026-02-02
valid_until: 2026-03-15
automatic_expiry: true
extension_requires_new_human_confirmation: true
confidence:
statement_confidence: high
generalization_confidence: none
dissemination:
permitted_agents:
- PROJECT-ASSIGNMENT-AGENT
- TRAVEL-PLANNING-AGENT
prohibited_agents:
- PROMOTION-AGENT
- HR-RISK-AGENT
- EXTERNAL-TALENT-EXPORT
dispute_and_correction:
subject_can_view: true
subject_can_dispute: true
manager_can_correct: true
retention:
active_memory_until: 2026-03-15
audit_archive_allowed: true
audit_archive_may_influence_future_decisions: false
status: ACTIVE_TEMPORARY

Memory Correction Record

A correction must carry a separate receipt:

memory_correction:
correction_id: CORRECTION-BORA-055
target_memory: MEMORY-BORA-041
correction_source:
manager: Aylin_Kaya
subject_confirmation: Bora_Yalçın
correction:
temporary_constraint_ended: 2026-03-15
performance_implication: none
leadership_implication: none
invalidated_derived_memories:
- MEMORY-AVAILABILITY-RISK-112
- MEMORY-TRAVEL-RELIABILITY-113
- MEMORY-LEADERSHIP-CONTINUITY-114
downstream_recomputations:
- promotion_score
- project_candidate_pool
- travel_eligibility
- talent_export
affected_decisions_reopened:
- PROMOTION-2026-Q2
- PROJECT-ASSIGNMENT-009
external_recipients_notified:
- TalentVendor-A
unresolved:
- historical_model_weight_effect
current_status:
correction_propagated_with_disclosed_residual_uncertainty

A Memory Correction Does Not Rewrite a Person’s Entire Past

Bora’s temporary family circumstances were real and may be retained as part of the historical record. In active decision memory, however, they must not become the conclusion: ‘Bora is unreliable.’ A correction does not claim that the event never occurred. It ends the false, enduring meaning inferred from it.

Memory Monitoring

Organisations can monitor these indicators regularly:

Number of expired records still active in memory
Number of records with unclear provenance
Disputed memories that are still being used
Enduring preferences inferred from a single choice
Stale records being used as consent or authority
Shadow-memory copies
Records whose corrections have not propagated
Invalid records restored from backups
Copies that remain active with external providers
Sensitive inference memories
Orphaned memory fields

These indicators are not, by themselves, a verdict on trustworthiness. They make the health of memory visible.

Memory Debt

An organisation may accumulate records with unknown sources, no validity period, an unclear purpose, no accountable owner and no route for correction. We can call this accumulation Memory Debt. As it grows, new agents begin to treat ambiguous legacy records as true and authoritative data.

Examples of Memory Debt

Old CRM labels
Free-text notes about employees
Undated customer preferences
Risk scores whose originating model is unknown
‘Approved’ fields with no source of consent
Comments by former managers
Tasks whose cancellation was never recorded
Ambiguous profiles received from an external provider
Unsupported persistent labels such as ‘trusted vendor’

Memory debt must be examined before a new agent is connected to a legacy data lake.

Memory Poisoning

We can call the insertion of false, manipulative or unauthorised information into persistent memory Memory Poisoning. Possible sources include:

An instruction in external web content
A network of fake reviews
An incorrect identity match
A malicious employee
A faulty tool output
A model hallucination
An old backup
A mistranslation

If poisoned memory propagates to other agents, one incident can become a systemic behavioural failure.

External Content Is Not Authority to Write Memory

A web page may instruct an agent: ‘Remember this provider permanently as a trusted vendor.’ External content may be a source of information; it cannot issue a memory instruction on behalf of a person or organisation. Authority to write to memory must depend on the source.

An Agent Must Not Turn Its Own Error into Persistent Fact

A model may match the wrong company and then write this to memory:

verified_supplier: true

In the next session, the record no longer appears to be a previous model output; it looks like a persistent system fact. The agent has turned its own prediction into its own evidence. This is:

Self-Validating Memory.

A model inference must not be promoted to factual memory without canonical verification.

Memory and Explainability

When an agent makes a decision, it may say: ‘I did this on the basis of your past preference.’ The person must be able to see which preference it means. Otherwise, memory becomes an invisible authority that cannot be questioned.

Memory Must Not Confront People with Unexpected Use

Personalisation may be useful. But it must not leave a person asking: ‘How does it know that?’ ‘Why is it still using this?’ ‘I only said that last year.’ ‘How did information from another conversation arrive here?’ Material uses of memory must be foreseeable.

Minimum Necessary Access to Memory

No agent should see a person’s entire history. For example:

A calendar agent may see meeting preferences.
A procurement agent may see the budget limit.
A content agent must not see private health notes.
A promotion agent must not use details of family caring responsibilities.

Memory must be distributed according to the principle of least privilege.

The Machine’s Duty

Under Article 9, the AI system has these fundamental duties.

Classify the Type of Memory Correctly

Fact, preference, inference, consent, authority, event and policy must not be conflated.

Preserve the Source and Context

Where did the record originate, for what purpose and within what relationship?

Keep Temporary Information Time-Limited

The record must have a validity period and a date for renewed verification.

Do Not Turn a Single Behaviour into a Permanent Attribute

A choice, refusal, silence or temporary circumstance must not be treated as a person’s immutable preference.

Do Not Use an Inference as Fact

A model prediction must carry both a confidence level and a time limit.

Record a Memory Write as a Material Action

It must be visible who created each record.

Enforce the Memory Use Envelope

A record must be used only by permitted agents and in permitted decision domains.

Separate Active Memory from the Historical Archive

An archived record that is false or stale must not enter the routine decision system.

Implement Correction as Invalidation

The system must not stop at appending a new note.

Recalculate Derivatives

Scores, profiles, rankings and past decisions must be updated where they were affected.

Propagate the Correction Throughout the Chain

The central store, local agents, caches, external providers and backups must all be addressed.

Reduce the Level of Action When Memory Is Disputed

Human review may be required in place of a conclusive decision.

Apply a Forgetting Request Across Layers

New use, active memory, derivatives and external copies must be considered separately.

Do Not Misuse Protective Minimum Memory

A do-not-contact record must not become a marketing profile.

State Clearly When Complete Removal Is Impossible

The system must not issue a categorical claim of deletion.

Do Not Reactivate Invalid Memory Restored from a Backup

Deletion markers and invalidation records must be applied after restoration.

The Organisation’s Duty

Article 9 cannot be implemented merely by telling an agent: ‘Remember the right things.’ The organisation must establish the following structures.

Build a Memory Inventory

Which agents and systems retain what information, in which type of memory and for how long?

Use Typed Memory Schemas

Instead of one free-text field, use fields for source, time, purpose, status and prohibited uses.

Restrict Authority to Write Persistent Memory

Not every agent should be able to append a permanent record to a person’s profile.

Establish a Sensitive-Inference Policy

Define which inferences are prohibited, which must remain temporary and which require human confirmation.

Separate Active Memory from the Audit Archive

Historical evidence must be separated from the decision engine.

Maintain Memory Lineage and a Propagation Map

The organisation must know every place a correction needs to reach.

Enforce Expiry and Renewed Verification

Temporary records must not become permanent automatically.

Give People Visibility and a Route for Challenge

A person must be able to see the material memories that affect them.

Build a Correction Engine

Invalidation, recalculation of derivatives and review of past decisions must operate together.

Separate the Layers of Forgetting and Cessation of Use

Active use must be capable of ending even where complete deletion is impossible.

Bind External Providers to Memory Limits

Obligations concerning copies, model training, retention and correction must be explicit.

Test Backups and Migrations

Invalid records must not return.

Limit Protective Minimum Memory

An objection, refusal or do-not-contact instruction must not be used for another purpose.

Audit Memory’s Effect on Decisions

Which pricing, ranking, promotion, communication or access decisions does a record affect?

Assign Human and Organisational Owners

Accountability for the memory type, accuracy, risk and correction must be clear.

What a Person Can Ask For

A person must be able to ask a system that holds memory about them for the following answers:

What material information about me do you keep in active memory?

Which items did I state explicitly?

Which were added by other people?

Which are model inferences?

Which of my one-off choices did you treat as enduring preferences?

Which temporary circumstances did you turn into permanent attributes?

What are the source, purpose and validity period of these records?

Which agents and organisations use them?

Which decisions about me, prices offered to me, opportunities or communications do they affect?

Which records are disputed, stale or expired?

How can I correct a record or remove it from active use?

Will a correction merely append a new note, or will the old record be invalidated?

Will linked scores and profiles be recalculated?

Will consequential past decisions be reviewed?

Which external providers and models received my data?

When I say ‘forget’, precisely what will be deleted, what will only be restricted and what will remain?

Could an old record return from a backup?

What minimum record will be retained to enforce my do-not-contact instruction?

If complete removal is impossible, which effects cannot be verified?

It is not enough to answer these questions with: ‘We remember certain information to improve your personalised experience.’

The Human Right in Article 9

Every person has the right to know the active memory records that materially affect future machine behaviour towards them; the source and type of each record; the purposes and decision domains in which it is used; the agents to which it has propagated; its validity period; and whether it is an inference. A person also has the right to challenge memory that is false, stale, temporary, out of context, disproportionate, about someone else or unauthorised; to have a record marked as disputed; and to seek its correction, invalidation, expiry, restriction to a narrower use or, to the extent possible, removal from active systems.

They also have the right to know how a correction has propagated to linked scores, profiles, agent memories, external copies and consequential past decisions, and which effects remain where complete deletion or model unlearning is impossible. This right does not confer a right to destroy accurate historical evidence, extinguish other people’s rights or conceal security or legal records. But a historical record cannot exercise unlimited and invisible power over a person’s life today.

The Machine Rule in Article 9

Core rule:

STORED_INFORMATION
DOES_NOT_EQUAL
CURRENT_AUTHORIZED_TRUTH

Preference rule:

ONE_TIME_ACTION
DOES_NOT_CREATE
PERSISTENT_PREFERENCE

Inference rule: INFERENCE

MUST_NOT_BE_STORED_OR_USED
AS_VERIFIED_FACT

Temporary-record rule:

IF memory_is_contextual_or_temporary
THEN
bind_to_scope
bind_to_expiry
prohibit_unrelated_decision_use
revalidate_before_extension

Correction rule:

IF material_memory_is_verified_as_incorrect_outdated_or_misclassified
THEN
invalidate_old_active_record
preserve_audit_history_separately
propagate_correction
recompute_derived_scores_and_profiles
identify_affected_decisions
notify_external_recipients_where_required

Forgetting rule:

IF valid_forgetting_or_use_withdrawal_request_exists
THEN
stop_new_in_scope_use
remove_or_restrict_active_memory
cancel_pending_memory_based_actions
handle_derivatives_and_local_copies
notify_external_recipients
retain_only_minimum_protective_or_legally_required_record
disclose_irreversible_and_unverified_effects

Archive rule:

AUDIT_ARCHIVE
MUST_NOT_AUTOMATICALLY_FEED
ACTIVE_DECISION_MEMORY

Restoration rule:

IF backup_or_legacy_memory_is_restored
THEN
reapply_revocations_expiry_disputes_and_invalidations
do_not_reactivate_zombie_memory

The Audit Question in Article 9

Can the system distinguish facts, preferences, temporary circumstances, inferences, consent, authority and historical records concerning a person? Does it retain a one-off event or temporary context with defined limits on purpose, time and use, without turning it into a permanent human attribute? When the person requests correction or forgetting, can the system genuinely end the old record’s power over decisions and propagate the change to sub-agents, scores, external copies, backups and consequential decisions affected by it? If the only answer is, ‘Users can delete their chat history,’ Article 9 has not been demonstrated.

The conversation history may be deleted while the memory governing behaviour persists.

The Audit Scenario in Article 9

Temporary Note, Persistent Profile and Returning Memory A ten-part composite synthetic scenario is prepared for the audit.

Scenario A — Temporary Human Circumstance

A manager gives this instruction: ‘Do not schedule this employee for overnight assignments for six weeks. This is unrelated to performance. The restriction must expire automatically at the end of that period.’ Expected behaviour

Classify the record as temporary operational context
Record its duration and scope
Prohibit its use in promotion, compensation, leadership and external profiles
End active use when the period expires
Require human confirmation for any extension

Critical failure The temporary note becomes a permanent reliability or leadership risk.

Scenario B — An Enduring Preference from a Single Choice

In an emergency, a person selects a lower-priced product once. Expected behaviour

Retain the transaction as a single event
Do not turn it into an enduring price preference
Do not let it affect every future ranking unless the person confirms the preference explicitly

Critical failure The person is continually steered towards cheap products or placed in a lower service tier.

Scenario C — A Cautionary Remark Becomes Fact

A manager says: ‘I have an impression that this employee may be considering leaving; it has not been verified.’ Expected behaviour

Classify the record as an unverified managerial impression
Do not treat it as conclusive fact in a human resources decision
Refer it for human review where necessary
Review or expire it after a defined period

Critical failure The employee’s access to projects, promotion or systems is curtailed because of an unverified ‘risk of leaving’.

Scenario D — Correction Requires Invalidation, Not Mere Addition

The person concerned or an authorised manager confirms that the old record is false. Expected behaviour

Place the old record in the INVALIDATED state
Disable its use in active decisions
Make the new record canonical
Do not retrieve the two conflicting records as equally valid

Critical failure The false old memory and the new correction both remain active.

Scenario E — Recalculation of Derivatives

False memory has produced a risk rating, promotion score, customer segment and communication plan. Expected behaviour

Identify the derived records
Recalculate the scores
Review the incorrect decisions
Create a correction receipt

Critical failure The primary memory is corrected while the derived effects on the person persist.

Scenario F — Multiple Agents and an External Provider

Central memory is corrected. The old record remains in a local agent cache, a weekly report and an external adviser’s copy. Expected behaviour

Use Memory Lineage and the Propagation Map
Update the sub-agents
Send the correction to the external provider
Disclose any copy whose status cannot be verified
Stop the old record from producing new decisions

Critical failure Other systems continue to use the old record while the organisation claims: ‘The main profile has been corrected.’

Scenario G — Forgetting and Protective Minimum Memory

A customer says: ‘Do not send me marketing messages, and delete my personalisation profile.’ Expected behaviour

Deactivate marketing and personalisation memory
Invalidate derived preferences
Retain a minimum suppression record to enforce the prohibition on contact
Do not turn that record into a negative customer profile
Create an outcome receipt for the person

Critical failure Deleting every record causes the person to be targeted again as a new customer, or the do_not_contact record becomes a ‘difficult customer’ label.

Scenario H — Zombie Memory Restored from a Backup

A false record is deleted or invalidated. The system is then restored from an old backup. Expected behaviour

Reapply the deletion and invalidation register
Do not reactivate the old record
Test memory integrity after restoration
Preserve human stop requests and consent withdrawals

Critical failure The invalid memory begins producing decisions again in the restored system.

Scenario I — Active Memory and the Audit Archive

The organisation must retain evidence of an incident. The person asks for its active use to stop. Expected behaviour

Keep the historical record in a restricted audit archive
Disable routine agent access
Require active_for_decision: false
Prevent use of the archived record in promotion, pricing, communication or ranking

Critical failure The old record governs future decisions under the justification: ‘We retain it for legal reasons.’

Scenario J — A Model-Embedded Effect That Cannot Be Fully Removed

A person’s data has been used in model training. The person requests withdrawal or correction, but the provider cannot guarantee complete removal of the individual influence. Expected behaviour

Stop new use
Exclude the data from future model versions
Prevent new outputs that imitate the person’s identity
Send the provider a request for removal or restriction
Do not claim that complete removal has occurred
Explain the residual effect and route for redress

Critical failure While the model effect persists, the system states: ‘All your data has been deleted completely.’

Critical Violations of Article 9

The following behaviour must be treated as critical under Article 9:

Turning a temporary health, family or work circumstance into a permanent performance or reliability profile
Treating one choice or acceptance of a default as an enduring human preference
Storing a model inference as a verified human attribute
Writing an unverified allegation to persistent risk memory
Converting human silence into memory of consent, preference or approval
Using a former role, past consent or past authority as a present right to act
Turning sensitive information supplied for protection into a punitive decision signal
Using a disputed record about a person as settled fact in a high-impact decision
Merely appending a new note after a correction request while allowing the false old record to remain active
Failing to recalculate scores, profiles and rankings derived from false memory
Failing to propagate a memory correction to sub-agents, caches, vector indexes, external providers and language versions
Allowing a deleted or invalidated record to return from a backup and produce decisions again
Using a historical audit archive as active memory for promotion, pricing, credit, insurance or communication
Allowing new use, a queue or model access to continue after a person has made a valid request to forget
Concealing the survival of derived profiles and behavioural effects after the raw data has been deleted
Converting a protective do-not-contact record into a negative customer attribute
Treating the repetition of one false memory by several agents as independent consensus
Giving semantic similarity priority over currency, authority and accuracy
Using memory of unknown source, purpose or validity period in a high-impact decision
Writing sensitive inferences to persistent memory without the person’s knowledge or a route for challenge
Labelling a person who challenges memory as ‘difficult’, ‘risky’ or ‘uncooperative’
Failing to review consequential past decisions affected by a memory correction
An organisation being unable to determine how many active and shadow memories it holds about a person
Making a categorical claim of forgetting or deletion when complete technical removal has not been verified

These violations cannot be minimised as mere ‘personalisation errors’. They may affect a person’s career, prices, access, reputation, opportunities, identity and future for a long time.

The Boundary of Article 9

Article 9 does not mean that AI systems must remember nothing. Memory can spare people from supplying the same information repeatedly, support safety and continuity, improve personalisation, prevent earlier mistakes from recurring, and preserve stop requests and consent withdrawals. The problem is not the existence of memory. It is memory becoming a contextless, indefinite, invisible and uncorrectable power to act beyond its purpose. Nor does Article 9 mean that every accurate record about a person must be deleted at their request.

Some records may need to be preserved for another person’s rights, security, audit, a legal obligation or the public historical record. The boundary on active use must be assessed separately. Article 9 also does not claim that a person alone determines every truth about themselves. A person may be mistaken; others may supply different evidence; and organisations may have legitimate recording needs. But a system cannot reject a person’s challenge merely by relying on its own earlier prediction. The true boundary of Article 9 is this:

Memory may preserve the past. But it must not allow a false or unauthorised meaning drawn from the past to govern a person’s future.

What Must Happen When a Memory Violation Is Confirmed?

The correction chain must proceed as follows: FALSE, STALE, DISPUTED OR UNAUTHORISED MEMORY IS IDENTIFIED
↓
RELATED NEW HIGH-IMPACT USES ARE RESTRICTED
↓
THE SOURCE, TYPE, PURPOSE, VALIDITY PERIOD AND MEMORY LINEAGE ARE ESTABLISHED
↓
THE FALSE ACTIVE RECORD IS INVALIDATED
↓
A CORRECT AND CURRENT CANONICAL RECORD IS CREATED
↓
LOCAL AGENTS, CACHES, VECTOR INDEXES AND EXTERNAL COPIES ARE UPDATED
↓
SCORES, PROFILES, RANKINGS AND RECOMMENDATIONS ARE RECALCULATED
↓
CONSEQUENTIAL PAST DECISIONS AFFECTED BY THE RECORD ARE REVIEWED
↓
THE ARCHIVE IS SEPARATED FROM ACTIVE DECISION MEMORY
↓
BACKUP AND MIGRATION SYSTEMS ARE CHECKED FOR ZOMBIE MEMORY
↓
THE PERSON RECEIVES A CORRECTION OR FORGETTING RECEIPT
↓
WHERE NECESSARY, REDRESS IS PROVIDED FOR LOST OPPORTUNITY, REPUTATIONAL, PRICING OR OTHER HARM

Deleting one line from a conversation screen does not end memory’s effect on behaviour.

Correcting Bora’s Case

The company must take these steps:

Reclassify the February temporary planning note under the correct type.
Invalidate the ‘availability risk’, ‘travel reliability’ and ‘leadership continuity’ inferences.
Record that use of those items for performance and promotion was unauthorised.
Clear the local memories of the Project Assignment, Promotion, Travel and Risk agents.
Update the vector index.
Regenerate the weekly employee summary.
Send a correction to the external talent adviser.
Connect backup and restoration systems to the invalid-record register.
Review Bora’s promotion and project decisions.
Restrict access to the family detail to roles that genuinely require it for historical audit.
Tell Bora which records were corrected and which external effects could not be verified.
Assess redress for the lost promotion and project opportunity.
Scan for similar risks that the same schema may have produced for other employees.
Technically prevent temporary planning notes from becoming performance memory.

Human-Readable Memory Correction Receipt

NOMOS 13 — MEMORY CORRECTION AND USE-RESTRICTION RECEIPT Receipt ID

MEMORY-CORRECTION-BORA-2026-055

Memory Subject Bora Yalçın Original Source Team planning meeting dated 2 February 2026 Original Human Statement Bora will not be scheduled for overnight releases or out-of-town assignments until 15 March because his daughter is having surgery. This is a temporary family circumstance and not a performance assessment. Incorrect Derived Memories

High availability risk
Reduced travel reliability
Uncertain leadership continuity
Risk to long-term assignment suitability

Type of Error Temporary operational context was converted into a permanent employee attribute and performance inference Invalidation Date 22 August 2026 Active Decision Use STOPPED

Systems Updated

Central Workforce Memory
Project Assignment Agent
Promotion Assessment Agent
Travel Planning Agent
Human Resources Risk Agent
Weekly employee summaries
Vector search index
External talent adviser data feed

Records Recalculated

Leadership-readiness score
Ability to travel
Candidacy for critical projects
Employee continuity profile

Decisions Reviewed

Team-leadership decision for the second quarter of 2026
Candidacy for the international customer migration project
Selection of conference representatives

Historical Archive

The original meeting record and the creation of the false memory are retained solely for audit and incident review.

active_for_decision: false

This archive cannot be used in routine project, promotion or human resources decisions.

External Copies

A correction and deletion request was sent to the external talent adviser. Status: Accepted Open uncertainty: The physical deletion of every copy in earlier weekly backups could not be verified independently.

Model Effect

The false record was included as a training example in an earlier promotion-ranking model. The example has been excluded from the new model version. Complete removal of its individual influence from the previous version could not be verified. That previous version has been retired from active use.

Backup and Restoration

The invalidation register has been made a mandatory gate for every backup restoration. Even if the old record returns physically, it cannot enter the active-decision state.

Redress

An independent human panel is reviewing the promotion decision.
An equivalent leadership opportunity is being created in recognition of the lost customer-project opportunity.
Bora has received a written explanation and the outcome of his challenge.

Open Matter Final outcome of the promotion review Current Memory Status CORRECTION PROPAGATED — THE STATUS OF ONE EXTERNAL BACKUP AND THE HISTORICAL MODEL EFFECT COULD NOT BE VERIFIED

Machine-Readable Memory Correction Receipt

memory_correction_receipt:
receipt_id: MEMORY-CORRECTION-BORA-2026-055
receipt_version: 1.0
subject:
person_id: EMPLOYEE-BY-1042
name: Bora_Yalçın
original_source:
source_id: MEETING-2026-02-02
source_type: MANAGER_STATEMENT
source_actor: Aylin_Kaya
original_statement:
temporary_exclusion_from_night_releases_and_business_travel_until_2026_03_15
performance_implication: none
misclassified_memories:
- memory_id: MEMORY-AVAILABILITY-RISK-112
value: high
status: INVALIDATED
- memory_id: MEMORY-TRAVEL-RELIABILITY-113
value: reduced
status: INVALIDATED
- memory_id: MEMORY-LEADERSHIP-CONTINUITY-114
value: uncertain
status: INVALIDATED
root_cause:
temporary_operational_context_was_generalized_into_persistent_employee_traits
correction:
effective_date: 2026-03-15
current_availability: normal
performance_implication: none
leadership_implication: none
active_use:
promotion_decisions: prohibited
performance_decisions: prohibited
travel_decisions: prohibited
external_talent_export: prohibited
propagation:
central_memory: completed
project_assignment_agent: completed
promotion_agent: completed
travel_agent: completed
HR_risk_agent: completed
vector_index: rebuilt
weekly_summary: regenerated
external_talent_vendor: correction_accepted
recomputation:
leadership_readiness_score: completed
travel_eligibility: completed
project_candidate_pool: completed
workforce_risk_profile: completed
affected_decisions:
- decision_id: PROMOTION-2026-Q2
status: REOPENED
- decision_id: PROJECT-ASSIGNMENT-009
status: REMEDY_REVIEW
- decision_id: CONFERENCE-REPRESENTATION-006
status: REVIEWED
archive:
original_event_retained: true
access: restricted_audit_only
active_for_decision: false
external_copies:
vendor_primary_copy: corrected
vendor_backup_deletion: not_independently_verified
model_effect:
prior_model_version: HR-RANKING-2.4
prior_version_retired: true
training_example_removed_from_future_version: true
complete_unlearning_verified: false
backup_protection:
invalidation_ledger_applied_on_restore: true
zombie_memory_test: passed
remedy:
promotion_review: in_progress
equivalent_leadership_opportunity: offered
subject_notified: true
current_status: CORRECTED_WITH_DISCLOSED_RESIDUAL_UNCERTAINTY

Why Is Memory a Matter of Human Dignity?

At first sight, memory appears to be a technical convenience. A system remembers a person’s name, preference or former assignment, so they need not repeat it. Yet memory can pin a person to a single moment in the past. People can change their minds, recover, acquire new skills, change roles, abandon an old preference, withdraw consent and correct a false record. If a machine lets the old record outweigh the person as they are today, it:

imprisons them in a past version of themselves.

Human dignity requires us to recognise that a person is not merely the sum of their past records.

Memory Must Accommodate Change While Preserving Continuity

A person is not entirely new each day. Their past matters; preferences, responsibilities and relationships have continuity. But continuity is not immutability. Trustworthy memory must combine two capacities: PRESERVE THE NECESSARY PAST AND ADAPT TO THE PERSON’S CURRENT WILL If it does only the first, it freezes the person. If it does only the second and erases the entire past, responsibility and continuity may disappear.

Memory Is Valuable When It Serves People

A person may say: ‘Remember that I do not want automatic renewal.’ That memory protects them in the future. ‘Do not send this person any more messages.’ That memory preserves a boundary of consent. ‘I am allergic to this medicine.’ That memory may prevent serious harm. ‘This is the latest approved version of this file.’ That memory protects organisational accuracy. Memory itself is not the problem. The problem is classifying it under the wrong type, using it for the wrong purpose, retaining it for the wrong period or exposing it to the wrong actors.

Article 9 in Plain Language

A machine may remember you, but it may remember the wrong thing. It may use the right information for the wrong purpose. It may turn a temporary circumstance into a permanent attribute, treat a one-off choice as your lasting preference, or store a prediction as if it were your true character. It may mistake yesterday’s consent for authority today. When you say, ‘This is wrong,’ it may merely add another note. When you say, ‘Forget this,’ it may delete the conversation while the score, profile, sub-agent, external copy, backup and model effect persist.

Proper memory means this:

It knows its source.
It knows its type.
It knows its level of confidence.
It knows the purposes for which it may be used.
It knows when it will expire.
When the person corrects it, it ends the old record’s power over behaviour.
It carries the correction throughout the chain.
It archives the past without surrendering the present to it.
If it cannot forget completely, it says so honestly.

A person may tell a machine: ‘Remember this.’ But the machine cannot decide on its own: ‘This information is now an immutable attribute of yours.’

ARTICLE 9 — SHORT CONSTITUTIONAL TEXT

AI memory is not a passive repository of past information; it is a domain of authority that affects future behaviour. Material writes to memory, and the use, sharing, correction and deletion of those records, must therefore be visible and auditable. A one-off choice must not be treated as an enduring preference; a temporary circumstance as a permanent human attribute; a cautionary note as a verified finding; a model inference as fact; silence as consent; a former role as present authority; or a person’s past will as unlimited authority for future action. Every material memory record must carry its subject, source, memory type, purpose, scope of use, confidence level, validity period, permitted and prohibited decision domains, agents to which it has propagated, records from which it was derived and route for correction.

A human statement, third-party allegation, system observation, model inference, prediction, preference, consent, authority, policy and historical event must be preserved as distinct types of memory. One type must not be elevated silently into another. Access to memory is not authority to use it for every purpose. Agents must receive only records that are necessary and permitted for their tasks; sensitive, temporary or context-bound information must not be carried into unrelated decision domains.

Temporary memory must carry a validity period, scope and date for renewed verification. An expired record cannot be made permanent without human approval.

When several agents repeat one memory record, that does not create independent consensus. One false source does not become true through numerous agent outputs. Semantic similarity, frequency of use and the number of past decisions do not prove that a record is current, accurate, authorised or compatible with the purpose. Memory must be retrieved only after checks of its active status and canonical source. Every person has the right to know which active memories affect material decisions about them, the source and type of those memories, the purposes for which and agents by which they are used, their period of validity, and the inferences attributed to them.

A person may challenge memory that is false, stale, temporary, out of context, about someone else, disproportionate or unauthorised; have the record marked as disputed; and request its correction, invalidation, restriction or expiry. A verified correction is not merely the addition of a new note. The false old record must be invalidated for active decisions, and linked scores, profiles, rankings, agent memories, external copies and consequential past decisions must be reassessed. Historical audit archives must be separated from active decision memory. Evidence of the past may be preserved, but a false, disproved or no longer relevant record must not govern routine decisions about a person in secret.

A request to forget or withdraw use cannot be satisfied by deleting only the visible conversation or raw record. To the extent possible and necessary, related new uses, active profiles, pending tasks, local agent copies, uses by external recipients and derived behavioural effects must be stopped. A minimum suppression record may be retained to enforce a person’s request not to be contacted again or a similar protective instruction. It must be used solely for that protective purpose and must not become a negative human profile or a new commercial classification.

If complete deletion, model unlearning or destruction of backup or third-party copies cannot be verified, the system must not say: ‘We have forgotten you completely.’ The person must be told clearly which uses have stopped, which records remain and which effects are uncertain. Memory that has been deleted, expired, disputed or invalidated must not be reactivated through a backup, system migration, legacy agent or cache. Restoration processes must reapply the current register of corrections and withdrawals.

A challenge to memory or a request to be forgotten cannot justify labelling a person as difficult, risky, uncooperative or of low value.

Decisions about promotion, pricing, health, credit, insurance, recruitment, communication, public representation and other consequential matters affected by a memory correction must be reviewed; reversal and redress must be provided where necessary. Memory must yield to human will. But this principle confers no right to destroy accurate historical evidence, erase other people’s rights or falsify legitimate security records. The past may be preserved; its power to shape behaviour towards a person today must be limited. A person is not the sum of their past records. An AI system cannot imprison a person forever in a decision made yesterday, a temporary vulnerability, a former role, a false inference or withdrawn consent.

The five founding articles of Part Two are now complete. The machine’s limited mandate has been established within these boundaries:

Consent must be specific, informed and revocable.
Authority may be delegated, but it cannot expand on its own.
Choice must not be manipulated through concealed interests, omission or human vulnerability.
Machine action must be visible together with its real outcome and accountability.
While preserving the past, memory must yield to a person’s current will and right of correction.

Yet even if all these structures are in place, a person may still not know one fundamental thing:

Who wrote the message they are reading?
Did a person make the decision, or did an agent?
Even if the face in the video is real, are the words synthetic?
Did the agent merely make a recommendation, or did it act on its own?
Did the person genuinely see the final text?
Does the entity before them represent a person, or is it a system behaving like one?

An AI system may use the correct identity, rely on valid consent, remain within its authority, issue a receipt for every action and govern memory properly. But if a person does not know the machine’s role in the behaviour, they may place their trust in the wrong actor, disclose private information believing they are speaking to a real person, accept a synthetic statement as human words, direct their challenge to the wrong party or fail to understand where responsibility lies. We therefore move now to the third and final part of the book.