Persist a fact, decision, lesson, or pattern so it can be recalled later by memory_search/memory_get or auto-injected into future sessions. Use for durable context (decisions, gotchas, preferences, recurring commands). For transient notes or duplicates, prefer memory_dedup_check first. Writes one SQLite row, queues an embedding (when enabled), and optionally creates/updates an Obsidian vault note. dedup="strict" refuses duplicates, "warn" stores with a warning, "off" bypasses.

Ranked full-text + decay-weighted search across SQLite, file-memory sources, and (when configured) the Obsidian vault. Read-only. Use the three-layer progressive-disclosure pattern: detail="index" (~30 tok/result, start here), "summary" (~80 tok), "full" (~150-300 tok, use sparingly). Follow-ups: memory_timeline(id) for chronological neighbours of one hit, memory_get(id) for one full body, memory_graph(id) to explore typed edges.

Fetch the full body, tags, and metadata of one memory by id (~300-800 tokens). Read-only. Use after memory_search(detail="index") to expand a single hit. Substrings inside <private> tags are redacted unless reveal_private=true (which emits an audit event). Returns not found if the id does not exist or has been soft-deleted.

Return the chronological neighbourhood of memories around an anchor id (~200 tok/neighbour). Read-only. Use after memory_search(detail="index") to recover work-session context for one hit. Cheaper than calling memory_get on each neighbour individually.

Cheap pre-flight (~50 tok/match): "would storing this content duplicate something I already have?" Read-only. Use before memory_store when overlap is likely, or before bulk imports. Returns up to limit existing memories with cosine similarity above the threshold (highest first). Computing the candidate embedding may make an outbound call to the configured provider (OpenAI, Ollama, etc.). If embeddings are disabled, returns a clear no-op message rather than silently passing.

Browse stored memories with optional filters — ordered by recency/importance. Read-only. Use to enumerate by type/scope/project. Prefer memory_search whenever you have keywords (relevance ranking). Use the same detail levels to control token cost.

Soft-delete a memory by id — the row is hidden from search/list/get but retained for audit. Idempotent. Use for accidental or obsolete memories. To replace one with a corrected version, prefer memory_store(supersedes_id=...) so history is linked.

Multi-action ADR log. Pick one via action: • store — record a new decision (title + body required; supersedes_id optional). Writes a row. • list — list recent decisions (read-only). • search — FTS over decisions (query required, read-only). Decisions outrank free-form memories by default and survive pruning longer. Use pitfalls_log for recurring problems.

Multi-action log of recurring problems and their resolutions. Pick one via action: • store — record a new pitfall (title + body required). Writes a row. • list — list open pitfalls (set include_resolved=true for all). Read-only. • resolve — mark a pitfall resolved (pitfall_id required). Use when something keeps biting and you want a queryable problem→resolution log. For one-off design choices, use decisions_log.

Reports utility rates of injected memories, token costs per search layer, auto-capture stats, compression activity, and prune suggestions. Read-only. Use to tune importance thresholds, find dead memories worth pruning, and verify that auto-capture/compression are earning their keep. Returns a no-op message when analytics are disabled.

Edit a memory in place — only the fields you pass are changed. Editable: title, content, tags, importance, memory_type, pinned. Use for typo fixes, retagging, importance tuning. For meaningful changes you'd want history for (e.g. a reversed decision), prefer memory_store(supersedes_id=...) instead. Side effect: when content changes, the embedding is re-queued and dedup may run. Returns not found for missing ids.

Toggle the pinned flag — pinned memories survive pruning and rank higher in search. Idempotent. Use sparingly: reserve pins for canonical decisions, user preferences, and high-leverage facts. Pinning everything defeats the purpose.

Force one compression cycle now: cluster similar memories by embedding similarity, merge each cluster into a canonical memory, and mark originals as compressed. Use after a bulk import or to sanity-check compression. Compression normally runs on the maintenance schedule. Side effects: writes merged memories, updates originals' compressed_into pointer, may call embeddings + LLM providers. Requires compression.enabled=true; otherwise returns a no-op message.

Export memories, decisions, and pitfalls as a portable JSON document. Read-only. Use before destructive maintenance (bulk delete, schema migration) so you have a clean restore point, or to transfer state to another memento-mcp instance via memory_import. Includes scope, tags, importance, and supersession links so round-trip preserves structure.

Import memories, decisions, and pitfalls from a JSON file produced by memory_export (server-local path, not a URL). Conflict handling via strategy: skip (default, safe to re-run) keeps existing rows on id collision; overwrite replaces them — use for authoritative restores. Side effects: inserts/updates rows in memories, decisions, pitfalls. Embeddings are queued asynchronously.

Create a typed directed edge between two memories. Edge types: relates_to, supersedes, caused_by, mitigated_by, references, implements. Use after memory_search to map dependencies, cause/effect, supersession chains, or references. Re-linking the same (from_id, to_id, edge_type) triple updates the weight (idempotent).

Remove a previously created edge — pass the exact (from_id, to_id, edge_type) triple used at creation time. No wildcards. Idempotent. Use to retract incorrect links. To retract a whole memory, prefer memory_delete (keeps audit trail).

BFS walk from one memory id outward, returning neighbour nodes + typed edges. Read-only. Use after memory_search to map related concepts, supersession chains, or the cluster around a decision. direction controls edge polarity (out/in/both); depth is capped at 5.

BFS shortest path between two memories — returns the chain of ids + edge types, or a no path message when unreachable within max_hops. Read-only. Use after memory_search (with both endpoints known) to trace cause-effect chains, supersession history, or dependency lineage.