@memory-river/adapter-mcp

MCP stdio server exposing Memory River to Claude Code, Codex, Cursor, Claude Desktop, and other MCP hosts.

Claude Code registration

Build the workspace, then add this exact entry to the MCP configuration:

{
  "mcpServers": {
    "memory-river": {
      "command": "node",
      "args": [
        "/path/to/memory-river/packages/adapter-mcp/dist/cli.js"
      ],
      "env": {
        "MEMORY_RIVER_DATA_DIR": "/home/you/.memory-river",
        "MEMORY_RIVER_RAM_DIR": "/home/you/.memory-river/ram",
        "MEMORY_RIVER_SESSION_KEY": "claude-code",
        "OLLAMA_URL": "http://localhost:11434"
      }
    }
  }
}

MEMORY_RIVER_DATA_DIR also accepts DATA_DIR; MEMORY_RIVER_RAM_DIR also accepts RAM_DIR. The embedding model defaults to hf.co/Qwen/Qwen3-Embedding-0.6B-GGUF and can be changed with MEMORY_RIVER_EMBEDDING_MODEL.

An OpenAI-compatible concentration LLM is optional. Configure it with MEMORY_RIVER_LLM_BASE_URL, MEMORY_RIVER_LLM_MODEL, and MEMORY_RIVER_LLM_API_KEY; the standard OPENAI_BASE_URL, OPENAI_MODEL, and OPENAI_API_KEY aliases also work. Without these variables, the server starts without an LLM. Recall, rehydrate, and store remain available; only LLM-dependent concentration is unavailable. Embedding-dependent operations still require the configured Ollama embedding service.

Surface and v1 boundary

The server exposes thirteen tools: memory_recall, memory_rehydrate, memory_archive, memory_store, memory_update, memory_set_status, gwm_on, gwm_off, gwm_status, gwm_update, skill_save, skill_load, and memory_river_info. memory_archive lets a host explicitly save conversation messages into the transcript store for later exact retrieval with memory_rehydrate. The server also exposes the memory_river_gap_aware MCP prompt, which tells the host agent how to judge lossy recall and rehydrate exact turns.

memory_update accepts a memory id plus one or more of text, category, importance, and metadata. memory_set_status accepts memoryId, toStatus (active, deprecated, superseded, or trashed), and optional supersededBy and meta. Setting trashed is the soft-delete operation; the MCP adapter does not expose physical deletion.

Version 1 is tools plus prompt only. It does not auto-archive host conversations or inject memory on every turn because MCP does not provide a standard host-fed conversation stream. Memory accumulates through explicit memory_store and memory_archive calls.

GWM (gwm_on/gwm_off/gwm_status/gwm_update) only stores task working-memory state through this adapter. Its automatic recall-biasing and drift detection run inside a host's context assembly, which MCP does not drive — so memory_recall does not yet apply GWM query expansion. Treat GWM as state-only until a host-neutral execution path exists.