Memory

Use durable memory as scoped evidence from an earlier moment, never as unquestioned current truth. Discover the stores actually available in the current runtime, retrieve only what the task needs, and keep all mutation approval-gated.

Workflow

1. Define the query, workspace, user, time horizon, and sensitivity boundary. Do not search unrelated projects or private connectors.
2. Discover available stores instead of assuming a vendor or path. Start with memory already loaded for the session, repo-local sources, and exposed read/search tools. Read references/store-adapters.md when selecting a store.
3. Retrieve narrowly with task-specific terms. Prefer indexes and targeted searches over loading whole stores. For an explicitly scoped local root, use python3 scripts/inventory_memory.py ROOT to inventory likely files without reading their contents.
4. Evaluate every relevant claim using references/memory-contract.md: record its source, date, scope, evidence kind, freshness, and conflicts. Separate stored evidence from present inference.
5. Verify drift-prone claims when verification is safe and proportionate. A current observation wins over an older memory; report the conflict rather than blending both into false certainty.
6. Answer with the minimum relevant context and citations or retrieval handles. State stale, conflicting, incomplete, or unverified evidence plainly.
7. Mutate only after an explicit user request and any runtime approval required for the action. Before promotion, read ../../../references/docs/promotion-matrix.md and route the item to its canonical owner.

Retrieval Contract