Associative recall
unimatrix_recall runs graph activation unless activate is false. unimatrix_search_memories does not run that stage. Multi-cue scoring combines activation, similarity, temporal cues, and memory strength.
Bring a decision, its source, the context around it, and its corrections into one governed memory layer. The 2.0 implementation adds associative dynamics to the existing retrieval and evidence pipeline.
unimatrix_search_memories combines semantic and keyword evidence. unimatrix_recall adds graph activation unless activate is false. Search does not run that stage.
unimatrix_recall runs graph activation unless activate is false. unimatrix_search_memories does not run that stage. Multi-cue scoring combines activation, similarity, temporal cues, and memory strength.
Librarian links indexed memories using similarity and temporal proximity. Co-activation can reinforce graph relationships through BCM / Hebbian updates. These are memory-ranking mechanisms, not training or fine-tuning a foundation model.
Worker jobs can replay, merge, and summarize related material into higher-level semantic memories, linking them to their sources. Adaptive utility and decay influence retrieval and maintenance; they do not guarantee that every fact will be remembered.
Veritas retains source lineage and supersession. Supporting and opposing evidence affect confidence signals. Unsettled conflicts can stay visible for review; a ranking score is not a probability that a claim is true.
The existing semantic and keyword path remains available. Compressed candidates, full-precision reranking, structured facts, and graph-aware fusion support different retrieval routes. A request does not automatically run every algorithm.
Memory distinguishes when information applied from when it was recorded. Explicit sessions can reinforce recent context. Timeline, provenance, and session tools depend on the tool list advertised by the connected endpoint.
Add sources and authority guidance, inspect the preview, and approve the handoff. Source-linked wiki views keep project context, decisions, sources, and contradictions together.
Normalize and chunk content, extract facts and relationships, classify memories, check contradictions, index for recall, and link the associative graph. Provider inference assists some stages; Librarian is the orchestration layer.
Customer clients connect to the web application through MCP at /api/mcp or the REST tools API. The web service delegates deeper processing to internal Fastify and worker services. PostgreSQL with pgvector is the canonical durable store; nested Spaces organize memories, with child Spaces presented as Locations.
Memory content is protected with AES-256-GCM at rest. Ingestion, classification, indexing, and recall can process plaintext; summaries, embeddings, and graph metadata are sensitive derived data. Row-level security and application checks are how tenant rows are scoped. They are not a formal guarantee that no bug can cross a boundary.
Where the endpoint advertises the upgraded tools, unimatrix_recall runs graph activation unless activate is false. unimatrix_get_context returns that result as current, contested, and superseded memories. Deployment requires the matching server, worker, database migrations, and internal service configuration.
// REST tool request body — POST /api/tools/call
{
"toolName": "unimatrix_get_context",
"args": {
"query": "Why did we choose idempotent retries?",
"intent": "implementation review",
"limit": 8
}
}unimatrix_feedback takes memory_id and outcome (helped, wrong, or stale). A local fallback reports activationRan: false and does not invent a context package. Inspect the endpoint’s advertised schemas for supported fields and limits. Authentication requires your own API key; never put it in a public repository.
| Layer | Current scope |
|---|---|
| 2.0 TypeScript server | AMG linking, explicit activation recall, co-activation reinforcement, interference handling, and worker consolidation are integrated in repository code. |
| Existing memory substrate | Hybrid recall, evidence-aware utility, source lineage, semantic chunking, temporal history, and session tools remain part of the implementation. |
| Rust engine and research | The Rust implementation and numerical conformance suite are separate engineering artifacts. Their presence does not mean hosted requests run on Rust or CXL hardware. |
| Hosted availability | This architecture describes the current code. Availability depends on the deployed revision, migrations, service health, plan, and tool surface. A code merge or passing math test alone does not establish production readiness. |
No universal accuracy, latency, storage reduction, or “remembers everything” guarantee is implied. Conformance tests check numerical behavior; retrieval quality and real client journeys require separate validation.