semantic-memory

Use claim-ledger to back material agent assertions with claim/evidence/provenance receipts. Create claims from facts, add evidence, verify integrity, and export bundles for audit.

Use context-governor for receipt-backed context compaction before risky handoffs or compaction events. Preserve high-risk context, store exact fallback receipts, and expand when omitted text matters.

Use when semantic-memory evidence must support an assertion, historical answer, correction, compaction, or action. Preserve witnessed retrieval, temporal scope, authority and native receipts. Applies to the RecursiveIntell memory stack; does not authorize memory maintenance or agent execution.

Explore the semantic-memory knowledge graph. Use when the user asks what is related to a topic, how two things or facts are connected, what is known about X across the graph, or wants to see clusters/communities in their memory. Chains hybrid search with second-order (discord) retrieval, shortest-path, and community detection, and can render the result as an interactive HTML graph.

Use semantic-memory-mcp LLM output parser tools to robustly parse JSON, string lists, choices, and numbers from raw LLM output without an additional LLM call. Handles think blocks, markdown fences, and malformed JSON.

Persist durable sourced facts to semantic memory when the user asks to remember a decision, preference, correction, or verified project fact. Excludes transient progress, secrets and unverified claims.

Audit and improve the semantic memory store. Use when the user asks to review, clean, audit, prune, or maintain memory, to find duplicates or contradictions, to check memory health, or to reconcile conflicting facts. Produces a health report (stats, contradictions, forgettable items, graph gaps), then — only with approval — reconciles via append/supersede and edge invalidation. Never destructively deletes.

Run semantic-memory store maintenance. Use when the user asks to compact, vacuum, rebuild indexes, re-embed, check integrity, or fix DB health. Covers the full maintenance workflow: integrity check, re-embed stale vectors, vacuum compact, reconcile FTS indexes, and verify search reproducibility via receipts.

Keep a codebase's memory current. Use when the user asks to sync, refresh, re-ingest, or update memory for a repository (e.g. after pulling changes or finishing work), or says "the memory for this project is stale". Re-ingests the repo idempotently (dedupe) so only new/changed components are added.

Run cargo fmt/clippy/test and store gate receipts before claiming work done.