Memory, Context & Knowledge
Three different information authorities: what is retained, what is relevant now, and what has been accepted as reusable truth.
Memory is retained information that may help later. Context is a task-specific package compiled from authorised sources for one execution. Knowledge is reviewed, accepted and reusable institutional information. Primitive deliberately keeps them separate so a remembered note cannot silently become canon and a useful context snippet cannot silently become authority.
Imagine a desk, a notebook and a reference manual. Context is what is laid out on the desk for this job. Memory is the notebook of things worth remembering. Knowledge is the reference manual that has been checked and accepted. The same sentence can move between them, but only through explicit steps.
Workspace now has five visible Memory lanes, candidate review, metadata for confidence/salience/retrieval strength, lifecycle maintenance, a Context Compiler that enforces scope and budgets, and a Knowledge boundary that admits accepted items only. Context compilation can draw from current objects, chats, files, compacted history, accepted memory, accepted knowledge, retrieval and prior runs while preserving public source provenance.
Retained is not relevant; relevant is not canonical
The five visible Memory lanes
| Lane | What it is for | Important behaviour |
|---|---|---|
| Short | Transient working notes | May fade/expire unless pinned or promoted |
| Long | Durable retained facts/context | Stays retrievable but is not automatically canonical Knowledge |
| Semantic | Meaning, definitions, entity relationships | Useful for conceptual retrieval |
| Episodic | What happened, in sequence | Preserves temporal/provenance-heavy recollection |
| Shared | Cross-user/agent visible memory | A visibility/ownership concern; sharing does not increase truth authority |
Candidate review prevents silent memory writes becoming truth
Workspace stores memory candidates separately from active memory. Candidates carry source kind/ref, content, memory kind, confidence, salience, provenance and review state. Accepted/merged candidates may become retained memory; rejected candidates remain evidence of what was considered.
A ContextPack is a bounded, authorised projection
The current compiler uses the actor's authorised scope, a context object and objective, then assembles canonical related context, recent conversation, extracted attachments, compacted history, retrieval results and prior successful runs. It enforces character/item budgets and removes revoked or inadmissible material rather than passing everything to the model.
| ContextPack field | What it contributes |
|---|---|
| objective | The bounded current task/question |
| context | The selected canonical object such as a task/chat/project |
| relatedContext | Direct canonical relationships, including accepted Knowledge only |
| conversation | Recent relevant messages |
| attachedFiles | Authorised extracted file text with integrity hashes |
| compactedHistory | Older chat epochs, explicitly non-authoritative narrative |
| retrieval | Ranked accepted Knowledge/Memory/files/messages/other authorised chunks |
| priorRuns | Relevant prior successful execution summaries |
| provenance | Query hash, route, versions, selected memory IDs and public source references |
| budget | What was included/excluded under the context limits |
Only accepted Knowledge enters canonical compilation
The compiler re-projects Knowledge through KnowledgeService and drops non-accepted items. Source revocation/unavailability is checked again at compile time. That means an old accepted statement can disappear from a future ContextPack if its required source is no longer valid or visible.
The same system can remember a disagreement without accepting either side
Primitive can retain both claims as Memory/Evidence, retrieve them into Context when relevant, and still keep Knowledge unresolved until review. This is essential for research, debugging, governance and long-running projects where yesterday's assumption may be today's superseded hypothesis.
Related pages
Canonical sources
These pages are explanatory read models. Where wording conflicts with implementation or a Control decision, the linked implementation/decision is authoritative and this page should be corrected.