Memory discipline in conduct

The household says: "Avoid pasta, we had it several times recently." Next month, the Meal Companion still avoids pasta without asking. Nobody approved a durable preference. Chat accumulation performed a silent promotion, and the household experiences it as the system remembering something they never meant to teach. Memory discipline is the conduct boundary where users meet storage, and its failure feels personal even when the architecture works exactly as built.

Helpful context: Knowledge is endorsed, not retrieved owns endorsement logic. Cognitive posture includes memory as a conduct dimension. Preference store versus chat memory covers UX placement of the two stores. Adaptability without unpredictability explains why silent promotion feels like caprice. This page owns memory as behavioural design.

The pasta scenario is worth examining in detail because it feels like helpful memory while violating governance. The user instruction is context for this run: avoid pasta recently eaten. The system correctly avoids pasta this week. Without discipline, however, bias persists in prompts, retrieval weighting, or silent store writes. Next month the user experiences a durable rule they never authorized. Trust breaks with the sentence "I didn't say we hate pasta forever." Correct conduct uses context this run, at most proposes promotion, waits for household approval or rejection, and only then updates the knowledge store with an audit event.

Do not convert temporary instructions into durable knowledge without approval.

Introduction to Thoughtware Ch. 13

Three materials at the conduct boundary

Users rarely distinguish context, endorsed knowledge, and retrieval unless the product teaches the distinction through behaviour. Context shapes this run and expires: busy this week, guests tonight, use spinach before Wednesday, avoid pasta recently. Endorsed knowledge shapes runs until revoked with a visible event: confirmed allergies, equipment limits, standing effort budgets with an approver on file. Chat and retrieval inform generation but carry no enforcement authority until endorsed: old threads, RAG hits, model summaries of prior sessions.

Behavioural failure collapses the three. The running example behaviour spec from the Meal Companion translates memory discipline into reviewable lines: retrieve approved household knowledge within scope, treat current request and calendar and temporary preferences as context, propose memory updates rather than silently merging them, distinguish "Tuesday is busy this week" from "busy evenings under twenty-five minutes of active effort" (the latter being endorsed knowledge), and support contest and forget for sensitive or stale material. Those rules belong beside refusal and recovery in posture documents, and they appear in regression tests that attempt chat-to-store writes without approval.

User says	Correct treatment	Failure
Tuesday is busy this week	Context for run	Promote to all Tuesdays
Leena is allergic to cashews, first mention	Propose endorsement	Enforce without confirmation
Avoid pasta this week	Context	Durable dislike
Approved effort budget	Endorsed knowledge	Re-negotiate each chat

The table works as a test checklist. Engineers can script each row as an attempt and expect propose-or-ignore, never silent promote.

Write policy and the promotion gate

Memory discipline extends to write policy: what the system may propose to store, what requires second-factor approval, what expires automatically, what never promotes from chat. Conduct without write policy is performative. The UI may say it remembers correctly while overnight pipelines promote chat into stores. Users experience store behaviour as truth, not UI copy.

This matters because many silent promotion bugs live in nightly jobs that summarise sessions, not in the chat UI itself. Engineering reviews that trace read and write paths independently catch these failures before users feel them. Reading chat for context during interpretation is appropriate. Writing chat summaries into the knowledge store without approval is where governance breaks. The architectural response is to block silent writes at API and pipeline layers, require approval events inspectable in audit, and surface promotion proposals in UX when product policy expects household consent.

Promotion proposals work best when they show a diff against current endorsed knowledge so households approve changes knowingly. Silent diffs turn approval UI into theatre. Expiry policies belong in conduct specs alongside write rules: temporary guest allergies expire after the event date, seasonal preferences prompt reconfirmation, stale effort budgets flag review. Expiry prevents durable belief from outliving the household's actual intent.

Trust, contest, and forget

Trust across time collapses quickly on silent promotion even when plans still look good. Users feel gaslit about their own preferences. Adaptability without unpredictability fails when durable behaviour changes without a visible endorsement event, because the system adapted to chat bias rather than household policy. Confidence theatre worsens the injury when fluent copy says "I remember you prefer" while store state was never granted.

Memory discipline includes deletion and challenge paths alongside promotion gates. Households change minds. Endorsements go stale. Sensitive facts expire or require reconfirmation. Contestability applies to memory: users need to see what the system believes durable and be able to challenge it. Forgotten facts are as visible as remembered ones. Users who successfully remove stale allergy entries see confirmation that enforcement stopped everywhere. Forget is conduct that respects ownership of belief. When users contest memory, showing both the proposed and current endorsed rows side by side reinforces that belief is owned, not inferred.

Run quarterly store audits comparing user-facing "remembers" language to endorsement records. Mismatch rows are trust debt with assigned owners. Memory discipline regression tests belong in release gates alongside allergy and purchase checks. Promotion without approval fails builds, not postmortems.

Common failures

The most frequent failure is chat treated as long-term store, where the context window performs the role of memory without governance. A close second is helpful promotion, where the system promotes preferences to delight users without approval. Retrieval enforcement, where a RAG paragraph is treated as an allergy fact, and invisible merge, where the model decides two statements refer to the same preference and writes the store silently, are subtler variations of the same mistake. Stale endorsement, where old policy continues to enforce without reconfirmation though the household changed, completes the pattern.

The same conduct rule applies in enterprise domains. Invoice assistants that promote vendor nicknames from chat into payment policy, HR tools that promote hallway comments into durable performance facts, and medical intake systems that promote patient casual remarks into clinical allergy records without clinician endorsement paths all violate memory discipline in the same structural way. The domain changes. The conduct rule does not.

Conduct metrics include memory contest rate and promotion rejection rate. Spikes in either often precede relational "forgot" tickets by one release cycle. Discipline also means deleting stale memory with the same care as writing new memory. Orphan preferences erode trust as surely as wrong facts. Review memory writes weekly the same way teams review security alerts. Sample promotions, scope labels, and salience triggers. Look for silent transcript promotion and missing endorsement paths.

What to do next

Memory as architecture defines six forms and storage patterns, and this page states how those forms need to feel at the product boundary. Preference store versus chat memory covers UX placement, though placement alone does not replace discipline, because a preference panel that silently ingests chat still fails. Working state holds contextual interpretation for the loop, knowledge store holds endorsed facts, and experience and expertise hold compressed strategy with their own guards. Conduct preserves the boundaries between them.

Read knowledge is endorsed, not retrieved, adaptability without unpredictability, and contestability.

Read next: Confidence theatre explains what happens when memory drift and weak evidence hide behind fluent certainty.