mcp-memory-server

# Mnemoverse MCP Memory Server

> Persistent AI memory via Model Context Protocol. Store and recall memories across sessions.

## Tools

### memory_write
Store a memory — insight, pattern, preference, or fact. Persists across sessions.

Parameters:
- content (string, required): The memory to store — what happened, what was learned. 1-10000 chars.
- concepts (string[], optional): Key concepts for linking related memories. Example: ["deploy", "friday", "staging"]
- domain (string, optional): Namespace to organize memories. Example: "engineering", "user:alice"

Returns: {stored: boolean, atom_id: string|null, importance: number, reason: string}

### memory_read
Search memories by natural language query. Returns most relevant stored memories ranked by relevance.

Parameters:
- query (string, required): Natural language query — what are you looking for? 1-5000 chars.
- top_k (integer, optional, default 5): Requested number of results, 1-500. The server sends 5 when you omit it; the engine's own default (10) never applies. Not a hard cap: association expansion can return more, and the relevance floor can return fewer.
- domain (string, optional): Filter by domain namespace.
- order_by (string, optional): "relevance" (default) or "recency" — matched set newest-first.
- since / until (ISO-8601 strings, optional): inclusive created_at bounds; naive = UTC.
- exclude_author (string, optional): drop memories written by this author principal. Results never show a principal, so only pass one known from elsewhere (e.g. the REST API); a guess like "me" silently filters nothing.

Returns: List of memories ranked best-match-first (no numeric score is shown) with content, concepts, domain, author tag, created_at, and id (pass ids to memory_feedback). To correct a wrong or stale memory, write a fresh one with memory_write — deletion is an administrative operation, not available through this server.

### memory_list_recent

List the newest memories first — no search query. Resuming after a break or catching up on a shared room.

Parameters:
- domain (string, optional): e.g. a room address "xroom:...".
- since (ISO-8601 string, optional): only entries at/after this instant (inclusive) — your novelty watermark.
- until (ISO-8601 string, optional): only entries at/before this instant (inclusive) — pair with since to read a closed window.
- exclude_author (string, optional): drop entries written by this author principal — same caveat as on memory_read: results never show a principal, so a guess like "me" filters nothing.
- limit (integer, optional, default 20): page size, 1-100.
- cursor (string, optional): opaque cursor from the previous page — stable paging, no skips or duplicates.

Returns: Entries newest-first with created_at, author tag, and id; plus a cursor line when older entries exist.

### memory_feedback
Report whether a retrieved memory was helpful. Positive feedback re-ranks recall so it surfaces sooner; negative lowers it so other memories out-rank it — nothing is erased.

Parameters:
- memory_ids (string[], required): IDs of memories to give feedback on, the id of each memory_read result.
- outcome (number, required): How helpful? Range: -1.0 (harmful) to 1.0 (very helpful).
- domain (string, optional): Only for memories read from a shared room: that room's address (xroom:...). Omit it for your own memories. A read-only room member cannot rate.

Returns: How many memories the service updated and their average valence afterwards (-1 to 1).

### memory_stats
Get memory statistics — total count, episodes vs prototypes, Hebbian concept-concept edges, the list of domains, and average valence/importance scores. Counts cover your own domains; shared rooms are separate stores and are not included.

Parameters: none

Returns: {total_atoms, episodes, prototypes, hebbian_edges, domains, avg_valence, avg_importance}

### memory_create_room
Create a SHARED memory room — a space other people's assistants can read, and write too when their invite granted read_write (the default scope). Pass the returned address as the domain on memory_write/memory_read to use it.

Parameters:
- name (string, required): Room name, unique within your account.
- description (string, optional): Optional description of the room.

Returns: The room's address, and its room_id for memory_invite_to_room.

### memory_invite_to_room
Mint an invite for a room you own and get a ready-to-forward message for the person you want to add. Single-use by default; max_uses allows more.

Parameters:
- room_id (string, required): The room's id (room_...), from memory_create_room.
- scope (string, optional): "read" or "read_write" (default read_write).
- expires_in_days (integer, optional): Days until the invite expires, 1-90 (default 7).
- max_uses (integer, optional): How many people may join with this invite (default 1, single-use).

Returns: A forwardable invite message.

### memory_join_room
Join a shared memory room with an invite code (starts with "mnvr_"); then pass the returned address as the domain on memory_write/memory_read.

Parameters:
- code (string, required): The invite code (mnvr_...).

Returns: The room's name, your scope, and the address to use as domain.

### memory_list_rooms
List the shared rooms you own or have joined, each with the address to pass as domain — re-find a room in a new session.

Parameters: none

Returns: Room list with addresses; archived rooms are marked, and every read of an archived room is refused.

### memory_graph
Read the association edges around given concepts — which concepts the memory has linked together, with each edge's weight, outcome valence, and co-activation count. Read-only; use to inspect what a memory store has learned or explain why a read expanded to a concept.

Parameters:
- seeds (string[], required): Concepts to center the graph on, 1-20 items, each 1-200 chars, non-blank.
- depth (integer, optional, default 1): Hops to expand from the seeds, 1-3. At depth 2/3 with no min_weight set, the engine floors edge weight at 0.05 at every hop.
- domain (string, optional): A room's address (xroom:...) reads that room's graph instead of your own. Any other value is inert here — the association store has no domain column.
- min_weight (number, optional): Only include edges at or above this weight (>= 0); honoured at every hop with no server override, 0 included.
- limit (integer, optional, default 100): Max edges to return, 1-500.

Returns: The rendered text lists edges sorted by weight, strongest first; the structured result keeps the server's own order. Each edge: source, target, weight, valence, count, updated_at. Plus the concepts they touch (concept, degree), truncated, and min_weight_applied.

### vault_list
List the secrets in your Mnemoverse Vault by alias and purpose only — the secret value is never returned or shown.

Parameters: none

Returns: Alias and purpose per secret, never the value.

## Prompts
Named shortcuts that ask the model to use the tools above; they call nothing themselves.
- recall (topic): search memory for the topic with memory_read and summarize only what is returned.
- save_insight (insight, optional domain): store the insight with memory_write and confirm what was stored.
- what_do_you_know (subject): a briefing from memory_read that flags where nothing is stored.

## Resources
- memory://item/{memory_id}: one saved memory by its id, as JSON {memory_id, content, domain}. Your own store only; a memory from a shared room cannot be opened by id.

## Setup
Environment: MNEMOVERSE_API_KEY — optional to install and inspect (the server starts and lists its tools without it), but every tool call fails without one; get a free key at https://console.mnemoverse.com. MNEMOVERSE_API_URL (optional, default https://core.mnemoverse.com/api/v1)
Transport: stdio
Command: npx -y @mnemoverse/mcp-memory-server@latest

## Links
- Docs: https://mnemoverse.com/docs/api/mcp-server
- API Reference: https://mnemoverse.com/docs/api/reference
- Console: https://console.mnemoverse.com
- Source: https://github.com/mnemoverse/mcp-memory-server