Check health of all mnemostack components.

Read-only, no side effects, no authentication required. Returns a JSON object with ok (bool) and per-component status for the embedding provider, Qdrant vector store, and optional Memgraph graph database. Use this to verify the memory backend is reachable before issuing recall queries.

Search indexed memories with hybrid recall.

Read-only, no side effects, no authentication required. Use this when you need raw memory matches rather than a synthesized answer. Returns a JSON object with ok, query, count, results, tokens_estimate (estimated total text tokens of the results), notes (the AUTHORITATIVE list of routine signals for stages that did not apply — e.g. temporal:no_parse on any query without a date; NOT a fault), and degraded (components that actually fell back, plus a deprecated back-compat duplicate of the routine tags until the next major; an entry absent from notes is a real fault). Results are ranked by reciprocal rank fusion of BM25, semantic, graph, and temporal retrievers when configured; each result includes id, text, score, sources, and payload. Stale facts (invalidated_at set) are hidden by default; use include_invalidated or as_of to see them.

Answer a question using retrieved memories.

Read-only, no side effects, no authentication required. Use this when you want a concise factual answer synthesized from memory search results instead of the raw matches returned by mnemostack_search. Returns a JSON object with ok, query, answer text, confidence (0.0-1.0), sources, notes (the AUTHORITATIVE routine signals — e.g. temporal:no_parse on any query without a date; NOT a fault), degraded (components that actually fell back, plus a deprecated back-compat duplicate of the routine tags until the next major), fallback_recommended, tokens_estimate (estimated text tokens of the context memories), tokens_used (LLM-provider-reported usage for the answer call; null when unreported), and error. Stale facts are hidden by default; use include_invalidated or as_of to see them.

Verify a citation: resolve a chunk id back to its source document.

Re-reads the CURRENT source and returns an honest verdict: intact / source_changed / moved (citation still supported), changed / missing (not supported by the current source), or unresolvable (cannot be verified from this process). Includes the snapshot-hash comparison and the fragment when locatable. Read-only; never mutates stored memory; runs outside the recall path. Resolution is confined to the corpus root recorded at ingest — there is deliberately no way for a caller to point it at another directory.

Mark memories stale by id, non-destructively.

A write tool (parallel to mnemostack_graph_add_triple). Sets invalidated_at (and optionally valid_until) on each point's payload without deleting or re-embedding it; invalidated facts drop out of default recall but stay reachable via include_invalidated / as_of. Points that do not exist are skipped. Pass index_root in multi-root collections to avoid marking another root's chunks stale. Returns ok, requested, and invalidated (the number of points actually updated).

Store a memory for the caller's tenant (the write counterpart of mnemostack_search).

A write tool: requires the write scope, embeds server-side, and stamps the process key's tenant so the memory lands in — and is recallable from — exactly this tenant's scope. Ids are deterministic from (source, offset, text): retries and repeated content return duplicate without a second embedding call. Long documents pass chunk=true and yield one result per chunk. Use mnemostack_invalidate to retract a stored memory.

Record explicit feedback for stateful recall learning.

Use signal='clicked' to also record inhibition-of-return exposure. Pass retriever labels from mnemostack_search results as sources so Q-learning can update source weights.