MGI-Mind MCP

Semantic (embedding) search across stored memories, ranked by relevance to query, with optional metadata filters. Read-only, no side effects. Returns up to limit results as text (library, score, id, content, author if set), or the literal 'No results.' when nothing matches. Use mind_browse instead for a query-less metadata listing (newest first, not ranked); mind_recall for error->fix procedur

Browse/list memories by metadata WITHOUT a search query: inventory, not ranking. Newest first. Use to answer 'what did agent X write recently', 'everything from this source', 'all memories since a date'. Same filters as mind_search.

Unified recall across all memory types at once (facts + memories + procedures) for one query. Current facts first, then semantic memories, then matching error->fix procedures. One call instead of three separate searches.

Decide whether to search memory BEFORE answering a user query. Returns priority (must-search / should-search / answer-directly), the reason, and which libraries to search first. Call this on a turn when unsure; it implements the search-before-answer trigger policy (named project, meta-cue like 'did I tell you', negation to verify, cross-session reference). Advisory: it cannot force a search.

Open the 3D memory visualization in the user's browser: the brain as glowing cores (memories, facts, regions) wired by neurons, with live pulses. Call this when the user asks to SEE or SHOW their memory / 'the brain' / how memory looks. Spawns a local viewer and returns the URL.

No parameter schema on file.

Store free-text content as a searchable memory. Content-addressed: the point id is derived from (library, content), so re-adding the exact same text upserts the same entry instead of duplicating it (and clears any archived or quarantined flag); any character change creates a new entry. Long content is split into overlapping chunks automatically. Fails if library does not exist (create it first w

Persist an externally-sourced snippet (code, doc, RFC quote, commit message, etc.) into mgi-mind with a mandatory provenance citation. The agent supplies the snippet AS PLAIN UTF-8: no HTML, no markup. Call this ONLY when the snippet was just produced by a code-search or doc-search MCP in the same session; do NOT fill provenance fields from memory.

Diagnose the local install: Qdrant reachable, embedding model and ONNX Runtime present, config valid. Read-only report by default. With fix=true it also DOWNLOADS missing pieces (Qdrant binary, embedding model, ONNX Runtime) to repair them; it never touches stored memories, facts, or the vault. Call this when a tool call fails with a connection/model/config error, before asking the user to reinsta

Build a text briefing for session start: operating rule, pinned mind_block entries, skill catalogue, last session summary, library list, and up to 20 current knowledge-graph facts. Side effect: every fact it surfaces is marked 'inherited', so a later mind_fact re-assertion of the same fact counts as reused context rather than an independent reconfirmation (discounted in the duel rule) until it is

No parameter schema on file.

Auto-ingest memory. PRIMARY: pass a candidates array of already-extracted items you judged worth keeping. BACKSTOP: omit candidates and pass raw text for marker-based heuristic extraction. Each candidate is secret-scrubbed and near-dup-checked before writing.

Record a procedural-memory lesson (error -> fix) for later mind_recall. Stored unverified by default (a manual lesson has no truth signal) and surfaced with low weight until confirmed; pass verified=true ONLY with a deterministic signal (test green / exit 0), or confirm it later with mind_outcome. Always inserts a new entry, never merges with an existing similar one. Returns the new procedure's id

Recall error->fix playbooks for an error that already happened, matched by error signature and/or task context. Read-only. Ranking blends relevance with trust: a verified or proven fix is boosted, a repeatedly-failing one demoted, but a strongly-matching unverified fix still surfaces. Returns up to limit playbooks as text, or a no-match message. Use mind_skill action="match" instead for a task y

Record whether a procedure returned by mind_recall worked when reused. worked=true/false bumps its success/fail counter and feeds its trust ranking, so the store self-corrects instead of ossifying on a bad fix; not idempotent, each call adds another data point. For anything other than a plain worked/did-not-work signal on a procedure (a typed signal like test_passed/code_compiled, a signal on a no

Record a typed external-signal outcome on any memory id, not only procedures: a test passed/failed, code compiled, a user confirmed/denied a fact, or a citing memory referenced this one. Idempotent on (memory_id, signal_type, source): re-posting the same triple updates the existing entry instead of appending a duplicate. Feeds the fact duel rule's external_signal_score. If memory_id is a procedure

List the most recently added memories (not facts, procedures, or skills), newest first by creation time. Read-only. Returns up to limit entries as text (library, content, source), or 'No memories yet.' when empty. Does not include a memory id; use mind_search or mind_browse when you need ids to act on.

Preview what mgimind consolidate would do: count of exact duplicates, near-duplicates, and cold (old + unused) entries. Always dry-run on the MCP surface; destructive consolidation stays on the CLI where the user types --apply explicitly. Use this when the user asks 'how much duplicate memory do I have?' or before suggesting they run the CLI command.

Restore an ARCHIVED (soft-forgotten) memory by id, returning it to search. Archiving (consolidate --archive-cold) reversibly hides cold memories instead of deleting them; this is the undo.

Fetch a webpage and convert it to Markdown via the external crw tool. Fails with an install hint if crw is missing or not on PATH, and fails if the page returns empty content. With save set, also stores the page as a memory (same rules as mind_add: save must already exist as a library, content is chunked, source is set to the URL) and returns a chunk-count confirmation instead of the page

Export memories to disk, one file per library. Formats json/md dump every stored memory (including archived and quarantined ones; facts, procedures, and skills are not included) grouped by library; format=instructions renders verified error->fix procedures only; format=profile renders a snapshot of pinned blocks + current facts + verified procedures + skills. Writes files under output, OVERWRITI

Reconcile a directory of Markdown files (Obsidian vault or plain folder) into a library: the files on disk WIN, so a file that differs from what is already stored replaces the stored version. Auto-creates library if missing. Defaults to a dry-run (apply=false) that only renders the plan (counts of new/replaced entries); pass apply=true to actually write. Idempotent: re-running with no file chang

Read-only summary: memory count per library and total, knowledge-graph fact count (plus a dependants/confidence distribution when facts exist), session file count, and whether the vault is initialized (never the secret count or values). No parameters. Use mind_consolidate instead to see duplicate/cold-memory counts.

No parameter schema on file.

Inspect, promote, or expire entries that the relevance gate filtered into the quarantine layer. Single tool with action: list (newest first, optional library filter), show (full content + gate reason by id), promote (the gate was too strict: move to ordinary memory by id), expire (the gate was right: delete by id; only ever touches quarantined points, never live memory, and stays recoverable fro

Vault is terminal-only by design. This tool explains how the user runs the equivalent mgimind vault command: secret values never cross the MCP channel. Single tool with action: store (instructions for storing a secret), get (instructions for retrieving), list (instructions for listing keys). Replaces mind_vault_store / _get / _list.

Manage the agent's session record across runs. Single tool with action: start (open a new session for an agent), last (read the previous session's summary), end (close the active session with a hand-written summary). Replaces mind_session_start / _last / _end.

Knowledge-graph facts (subject -> predicate -> object). Single tool with action: add (insert a new fact; returns verdict=invalidated and refuses if this exact triple was explicitly invalidated before - call action=restore first), query (text-matches subject against subject/predicate/object, returns current facts as text, or the superseded history with history=true, or a point-in-time snapshot

Library namespaces. Single tool with action: create (new library by name), list (all libraries with counts), delete (remove a specific memory by id within a library: destructive). Replaces mind_create / mind_list / mind_delete.

Register and inspect predicate cardinality (single | temporal-single | multi), which controls how mind_fact action="add" detects conflicts: Multi never conflicts; Single/TemporalSingle flag a conflict when a second distinct object is asserted for the same subject+predicate. Unregistered predicates default to Multi. Call action=register once per predicate before relying on conflict detection for it

Pinned core-memory blocks (persona / user / current-project): a few small, named notes that are ALWAYS injected at the top of the session context, not ranked retrieval. Use for stable, always-relevant facts the agent must not have to search for. action=set overwrites, get/list read, remove deletes. Content is capped (4KB/block, 32 blocks); this is core memory, not a second store: put searchable kn

Skills: the house way of doing a kind of work, matched against the task you are ABOUT TO START (mind_recall is the other half: it answers an error that already happened). action=match takes a task description and returns the playbooks that apply, ranked by relevance and by how they have actually worked out. Call it before starting non-trivial work, the way you would check for a house rule. action=