An MCP server that exposes a MemoWeft memory database as Model Context Protocol tools, so an external AI client (Claude Desktop, Cursor, or any MCP-capable host) can recall stored knowledge and record verbatim user messages.

This is an external integration package. It wraps MemoWeft's public Core facade (createMemoWeftCore) and does not re-implement any memory logic. It surfaces a deliberately narrow, mostly read-only tool surface.

Tools

Read (5)

Tool	Core method	What it does
`memoweft_recall`	`core.recall`	Recall cognitions relevant to a query (with confidence + credibility).
`memoweft_list_cognitions`	`core.memory.listCognitions`	List all cognitions for a subject, with evidence links and effective confidence.
`memoweft_list_evidence`	`core.memory.listEvidence`	List all raw evidence (sources) for a subject.
`memoweft_list_events`	`core.memory.listEvents`	List all events for a subject, with the evidence ids each covers.
`memoweft_graph`	`core.graph.buildMemoryGraph`	Build a memory graph payload (nodes, edges, stats).

Write (1, light)

Tool	Core method	What it does
`memoweft_ingest_user_message`	`core.ingestUserMessage`	Store a single verbatim user message as spoken evidence. Does not update the profile, run consolidation, or grant any cloud-read authorization.

SECURITY — this is an autonomous external call surface

MCP tools can be invoked autonomously by an external AI client, without a human approving each call. The tool surface here is chosen accordingly.

5 tools are read-only. They never mutate memory.
1 tool is a light write (memoweft_ingest_user_message): it records one raw user message as spoken evidence. It does not touch the profile, run consolidation, or change any authorization bit. observed evidence and cloud-read defaults are untouched.
Destructive and authorization-changing operations are intentionally NOT exposed as tools. The following MemoWeft Core methods are never registered:
- invalidateCognition, removeEvidenceSafely, removeCognitionSafely, mergeCognition, archiveCognition, resetSubject — destructive / data loss.
- updateEvidenceAuthorization — changes the cloud-read authorization bit (privacy-sensitive).
- handleConversationTurn, updateProfile — run the full ingestion/consolidation pipeline and rewrite the profile.
- ingestObservation, portable.* (export/import) — bulk data movement / observation authorization surface.
Handing any of these to an autonomous caller carries too much risk. Managing memory (deleting, merging, changing authorization, resetting) stays a human-supervised operation in the host app, not an autonomous tool.
Tool descriptions use neutral protocol wording. MemoWeft Core is headless — no persona, no anthropomorphized "I remember things about you" copy.

Install

npm install @memoweft/mcp-server

memoweft is a peer dependency (^0.5.0); install it alongside if it is not already present:

npm install memoweft@^0.5.0

Run

The server talks MCP over stdio. It is normally launched as a subprocess by an MCP client. It requires MEMOWEFT_DB_PATH — the path to your MemoWeft database file (there is no default path; the server refuses to guess one). Use :memory: for an ephemeral database.

Example MCP client config (e.g. claude_desktop_config.json):

{
  "mcpServers": {
    "memoweft": {
      "command": "npx",
      "args": ["-y", "@memoweft/mcp-server"],
      "env": {
        "MEMOWEFT_DB_PATH": "/absolute/path/to/memoweft.db"
      }
    }
  }
}

Model and embedding configuration follow MemoWeft's standard .env loading (see the main package). If they are missing, the server still starts; tools that need a model degrade gracefully (core.health() reports readiness) rather than crashing.

Programmatic use

import { createMemoWeftCore } from 'memoweft';
import { createMcpServer } from '@memoweft/mcp-server';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';

const core = createMemoWeftCore({ dbPath: process.env.MEMOWEFT_DB_PATH! });
const server = createMcpServer(core);
await server.connect(new StdioServerTransport());

Registry

server.json holds the MCP registry metadata. The name io.github.memoweft/memoweft is a placeholder GitHub namespace. The registry namespace ownership (which GitHub account/org backs io.github.memoweft/*) and the actual publish step are done manually by the maintainer (this machine has no gh installed). Publishing order: publish memoweft@0.5.0, then @memoweft/mcp-server to npm, then submit to the registry with the mcp-publisher CLI. package.json's mcpName must stay byte-for-byte equal to server.json's name.