Memory

A household opens the dinner planner on Sunday evening. The screen shows Tuesday marked busy, a cashew allergy enforced on every candidate meal, and a note that last month a similar week failed when two heavy recipes landed on short evenings. Three facts feel like one kind of "memory." Architecturally they are three different forms with different lifetimes, approval paths, retrieval rules, and UX obligations.

Product teams often collapse everything into a single context window or vector index. The demo remembers. Production surveils, drifts, and lies with fluent verbs. Users hear that the system "knows" when it retrieved chat. Auditors cannot trace what was endorsed versus what was mentioned once. Memory becomes trustworthy when each record carries a type and governance that matches how the product behaves.

Helpful context: Six forms of memory teaches discrimination drills. Knowledge is endorsed, not retrieved owns endorsement logic for durable belief. This page is the canonical overview of memory forms in Thoughtware.

Why one box fails

When memory is undifferentiated, three failures compound. Surveillance UX emerges because the product cannot explain what is temporary versus standing policy, and everything feels remembered forever. Unstable learning follows because experience, chat, and endorsed knowledge merge in retrieval, so the system appears to learn when it is mixing materials. Authority collapse arrives because temporary corrections silently harden into belief without approval events anyone can audit.

The Weekly Meal Companion grounding scenario makes the distinction teachable. "Tuesday is busy this week" is context, true for this run and expected to expire. "Busy evenings stay under twenty-five minutes of active effort" is endorsed knowledge, a household policy with approval and enforcement. "Last month we repaired after placing two heavy meals on busy nights" is experience, a selected episode that may inform retrieval but remains a belief-candidate until promoted. "When at least two evenings are low-time, anchor easy meals first" is expertise, a versioned pattern with applicability guards rather than a vibe retrieved from chat.

Distinguishing those four sentences is the discrimination exercise that prevents every downstream confusion. "Avoid pasta because we had it recently" is a temporary correction, context for this run. It must not silently become durable dislike without approval. That single mistake is how conversational products convert negotiation into ungovernable policy.

Memory becomes useful when the system knows what kind of memory each thing is.

Introduction to Thoughtware · Ch. 23

Six forms at a glance

Form	Records	Typical lifetime
Context	This situation, this request	Run
Working state	Goal, strategy, progress, gaps	Run (structured)
Knowledge	What may be believed	Durable until revoked
Experience	What happened	Selected episodes
Expertise	What tends to work	Versioned primitives
Evaluation memory	Regression, calibration	Release discipline

Working state is orchestration memory: the fields an agent loop reads and updates, distinct from dialogue transcript. Transcript is an audit record for humans. Working state is what pursuit consumes. See working state versus transcript for the full treatment of that boundary.

Evaluation memory protects standards across releases. It belongs in QA and release gates, injected into generation runs through explicit harness design rather than mixed with household preferences in ordinary retrieval. The distinction between one memory box and typed memory determines whether belief stays proportionate. A single undifferentiated store mixes chat, facts, and benchmarks, making learning unstable and authority invisible. Typed memory gives each form its own retrieval, approval, and forgetting rules, so the system can answer why it believes what it believes.

What this looks like in the household planner

In the household dinner planner from the Thoughtware system, memory forms appear together but require separate architectural treatment.

Working state holds the pursuit record: interpreted week, candidate meals, current plan version, accepted meals, unresolved gaps, retrieved expertise references, and loop budget. The Meal Planning Agent reads these fields each cycle. They version when the household accepts Monday through Wednesday or when a local repair patches Thursday.

Knowledge store holds endorsed facts: Leena's confirmed cashew allergy, household mushroom dislike after approval, the twenty-five-minute busy-evening rule, and the constraint that grocery purchase requires human approval. These records carry provenance and scope. Allergy enforcement runs in deterministic code against endorsed knowledge, not against whatever the model retrieved from chat.

Expertise library holds Busy Week Pattern v0.2 with an applicability guard: at least two low-time evenings, busy evenings clearly identified, no unfamiliar medical restriction, leftovers acceptable, no conflict with temporary request. The agent retrieves expertise only when the guard passes, and otherwise full deliberation returns. See applicability conditions.

Context assembles per run: current calendar, guest count, spinach deadline, temporary "avoid pasta" instruction. Context informs generation, but it does not override knowledge without an explicit promotion path. The four stores sit alongside each other in the architecture, and their separation is what makes each one governable.

Governance, conduct, and promotion

Memory typing is architectural, but users experience it as conduct. When product copy says "remembers," engineers map the claim to a form. Endorsed knowledge changes through visible approval. Context expires. Experience may suggest but does not create belief. Expertise runs behind guards. Violations are trust incidents even when output looks fluent.

Grounded verbs map everyday language to these forms: "knows" maps to endorsed knowledge, "remembers (for this week)" maps to context or working state, and "learned" maps to experience or expertise promotion with evidence rather than retrieval alone. These verb distinctions carry into preference store versus chat for standing choices versus ephemeral negotiation, and into memory is part of UX for surfacing and editability.

Memory sits inside the Thoughtware system alongside libraries, evaluation, deterministic shell, and human authority. The ecosystem layers of retrospection, bridging, and approval govern how experience becomes knowledge and expertise without silent drift. Libraries package judgment. Memory holds what the organisation and user have agreed the system may treat as true, situational, or historical. When teams sketch architecture, memory forms belong on the same diagram as cognitive units and agents, not as a database afterthought. What persists, who approves persistence, and what retrieval may inject into a bounded judgment are design decisions with safety consequences.

Experience and context do not become knowledge or expertise by accident. Each promotion has a distinct path with different evidence requirements. Experience to knowledge happens when a household or organisation decides a repeated pattern deserves standing policy. The path runs through bridging and approval: scope is named, approver recorded, enforcement updated in deterministic code where the fact is exact enough to close. Experience to expertise happens when a trajectory pattern earns reuse under guards. Busy Week Pattern v0.2 exists because eval showed compression worked on a defined class of weeks, failure modes were documented, and fallback returned full deliberation when guards failed. Expertise without guards is shortcut theatre.

Context to knowledge is the most dangerous silent path. "Avoid pasta this week" must expire or pass through explicit "save as preference" UX. Any pipeline that writes durable stores from run context without approval is a severity-one defect, because that pipeline converts negotiation into ungovernable standing policy. Knowledge revocation matters as much as promotion. When mushroom dislike is reversed, supersession belongs in the preference store and audit log rather than buried in embedding space.

Common mistakes

The most frequent failure is chat treated as a knowledge store, where long context windows become durable belief and a cashew mention in passing becomes "known." A close relative is silent promotion, where temporary instructions merge into embeddings without approval events. These two mistakes produce a system that cannot distinguish standing commitments from conversational artifacts.

Evaluation artifacts leaking into user retrieval is subtler but equally damaging, because regression sets bias generation invisibly when mixed into ordinary context windows. Using transcript as working state is another common failure: accepted meals live only in message history, local repair loses footing, and eval cannot score transitions on structured fields. Running one retrieval index for everything destroys governance distinctions, because similarity search across forms cannot respect the authority boundaries that keep memory honest.

When reviewing a memory design, the productive questions concern form: which form is each field, who may write it and under what approval, what retrieval may inject it into a cognitive unit call, what happens at run end or account merge or model upgrade, and whether product copy using "knows" or "remembers" can point to an inspectable record. If any answer is "the model will figure it out," the design is undifferentiated memory wearing a typed label.

What to do next

Every durable record deserves a form label before storage technology enters the conversation, because the form determines retrieval rules, approval paths, and forgetting rights that the database alone cannot enforce. Retrieval, bridging, approval, and merge for knowledge promotion paths stay separate from each other and from context assembly. Temporary corrections never promote without visible approval and scope. Teams that drill discrimination with six forms of memory before schema design consistently avoid the undifferentiated-store mistakes that surface months later as trust incidents.

See grounded verbs for the verb-to-form map and memory discipline in conduct for behavioural rules at the boundary. Read episode selection for how experience enters retrieval, and retrospection, bridging, and approval for ecosystem paths that connect forms over time.

Read next: Six forms of memory.