Six forms of memory

Product review opens a screenshot: "The system remembered we hate mushrooms." Engineering checks the logs. Mushrooms appeared once in chat, never approved, mixed into retrieval alongside endorsed allergy rules. Support escalates because the household thought standing policy changed. Auditors find no approval event. The sentence conflated context, chat, and knowledge, three forms with incompatible governance treated as one.

This failure is predictable when teams skip discrimination. Storage technology arrives before memory architecture. Vector indexes ingest everything. UX copy uses "remembers" without specifying what kind of memory is claimed. The fix is naming the form before naming the database.

This page teaches discrimination: which form a piece of information belongs to, so storage, UX, conduct, and evaluation stay aligned. Memory as architecture is the primer overview. This page is the reference table and drill set for schema review and copy audits.

Reference table for the household planner

The Weekly Meal Companion grounding scenario supplies concrete contents for each form. Every new field in an architecture review receives a form label before implementation.

Form	Example contents	Lifetime	Governance
Context	Current request, calendar, guests, spinach deadline, temporary "avoid pasta"	This run	Assembled per pursuit
Working state	Goal, interpreted week, candidates, plan version, gaps, checkpoints, budget	This run (structured)	Agent loop reads/writes
Knowledge	Cashew allergy, mushroom dislike, 25-minute busy rule, purchase needs approval	Durable	Endorsement + scope
Experience	Week 14 trajectory: two heavy meals on busy nights, repair after objection	Selected episodes	Capture policy
Expertise	Busy Week Pattern v0.2: guard, skeleton, failure modes, fallback	Versioned	Eval + promotion
Transcript	User messages and assistant replies in order	Audit	Not pursuit state
Evaluation memory	Busy-week regression set, grader calibration cases	Release discipline	QA owns

Evaluation memory is the sixth durable architectural form in the books: artifacts that protect standards across releases. Transcript is the dialogue record people confuse with working state. It is audit material, not orchestration state.

Each form exists because it answers a distinct question. Context answers what is true about this request and this week. It expires when the run ends or when assumptions materially change. Treating context as knowledge produces surveillance and silent policy drift. Working state answers what the agent currently believes matters for pursuing the goal. It is structured, versioned, and machine-readable. Fields like accepted_meals[] and unresolved_gaps[] enable local repair, escalation, and trajectory evaluation. Knowledge answers what the system may treat as believed until revoked. Endorsement, scope, and provenance are mandatory. See knowledge is endorsed, not retrieved.

Experience answers what happened in selected past trajectories that might inform interpretation. Experience proposes. It does not create belief without promotion. Episode selection policy matters here, and episode selection covers retrieval discipline. Expertise answers what pattern has earned reuse under guards. Expertise is versioned, evaluated, and applicability-checked, not "the model got faster at this." Transcript answers what was said, in order, for humans and audit. Loops that parse transcript to recover acceptance are using the wrong source, because that information belongs in working state.

Discrimination in practice

The discrimination drill is simple: given a sentence, name its form before discussing storage or UX.

Sentence	Form	Why
"Parents staying this week, plan for six."	Context	True for this run, expires after
"Accepted meals: Mon Wed Thu, open gap: Friday variety."	Working state	Structured pursuit progress
"Leena: confirmed cashew allergy."	Knowledge	Endorsed fact with approval
"Similar request three weeks ago required local repair on Thursday."	Experience	Retrieval candidate, not belief
"Reuse busy-week anchoring when guard passes."	Expertise	Promoted pattern with applicability
"User: 'Can we do something lighter Tuesday?'"	Transcript	Dialogue for audit

The mistake drill is equally revealing. Treating transcript as working state means loops cannot repair and eval cannot score transitions. Promoting "avoid pasta this week" to knowledge is an endorsement violation. Injecting evaluation memory into generation turns benchmarks into hidden policy. Calling expertise retrieval "remembering" creates false familiarity when the guard gets skipped. Storing context in the preference store turns temporary negotiation into standing rule.

These drills work in sprint planning when someone says "the system remembers X." The answer is always a form label first. Labels applied in schema design, where every new persistence field gets a form tag in the ticket, prevent months of undifferentiated storage. Labels applied in copy review, where marketing and support drafts that use "remembers" or "knows" require a form citation, prevent trust mismatches between what the product claims and what the architecture delivers. Labels applied in incident retros, starting with "which form was confused?" before "which model version ran?", reveal that most memory incidents are taxonomy failures rather than temperature failures. New engineers receiving ten sentences from the household planner scenario and labelling forms under time pressure in a fifteen-minute drill prevents the confusion that compounds silently through later design.

Composite scenarios stress-test labels further. When a user says "parents visiting" mid-run, that is a context update, a working state refresh, and a transcript entry, but not knowledge. When a household approves "no shellfish ever" in settings, that is knowledge in the preference store, an approved source, with shell enforcement updated. When the agent retrieves a week-14 failure narrative, that is an experience candidate for CritiquePlan, not belief and not expertise until the promotion path completes. When Busy Week Pattern v0.2 runs after guard checks, that is expertise retrieval with version pin, eval-backed, fallback documented. If the team cannot label all four in under five minutes, the taxonomy is not yet muscle memory.

Working state, transcript, and evaluation boundaries

	Working state	Transcript
Purpose	Pursue and repair goal	Record dialogue
Shape	Versioned fields	Ordered messages
Consumer	Agent loop, evaluators	Humans, audit
Example field	accepted_meals[]	"Looks good for Mon-Wed"

Full treatment lives in working state versus transcript. The distinction is among the highest-leverage fixes for agent reliability, because loops that consume structured state can patch locally, escalate with context, and expose transitions to eval. Loops that parse transcript lose accepted work across interruptions.

Transcript supports audit and dispute resolution. Evaluation memory supports release discipline and grader calibration. Neither belongs in ordinary retrieval paths that fill cognitive unit holes during household planning. Mixing these forms into retrieval pipelines without discrimination produces outputs that optimize for benchmark shapes or narrative coherence rather than for the household's actual goal. When eval artifacts leak into generation, the system optimizes for benchmark shapes users never agreed to. When transcript summarisation replaces working state, the system optimizes for narrative fluency over goal pursuit. Boundaries between these forms are explicit in data flow diagrams attached to the map.

Forms connect to sources and the ecosystem

When a cognitive unit fills an information hole, the form tells you lifetime and governance. Four kinds of source tells you provenance: approved, system, observed, or unaccountable. A fact can be observed this run (context) or approved for durable use (knowledge). Mixing forms without labels hides the weakest link in composed pipelines.

Memory forms describe lifetime and governance. Source kinds describe provenance at input time. A field can be knowledge (form) filled from an approved source (provenance). Conflating the two lenses loses information: endorsed knowledge can still arrive through a broken pipeline as unaccountable retrieval if source kinds are not enforced at cognitive unit boundaries. Pair form labels with source classification when cognitive units consume external facts.

Memory stores live in the ecosystem surrounding agents and Skills: preference services, episode archives, expertise catalogs, eval artifact repos. Agents orchestrate. Typed stores ground what retrieval may inject. Product language reflects forms even when users never see the taxonomy. "Household preference" implies endorsed knowledge with edit surfaces. "For this week only" implies context. "Based on how last month went" implies experience retrieval, not belief. Mislabelled copy trains users to trust the wrong material. Preference store versus chat specifies the UX pattern for standing choices versus ephemeral chat. Memory discipline in conduct states behavioural rules at the boundary.

What to do next

Form labels belong in data models, APIs, and event schemas before any persistence field merges without a typed column or metadata. Silent promotion from chat to knowledge in ingestion pipelines is the defect that converts negotiation into ungovernable policy, and blocking it early is cheaper than repairing trust later. Architecture reviews that reject unlabelled persistence, copy audits that match marketing screenshots against endorsement reality, and cognitive unit specs that pair form labels with source classification at boundaries all reinforce the same discipline from different angles.

See memory as architecture for the primer narrative and the Thoughtware system for how memory forms sit among agents, libraries, and shell code.

Read next: Four kinds of source.