Skip to content
MCP n24q02m logoMCP n24q02m logoMCP n24q02m

Overview

Mnemo MCP – Overview

Section titled “Mnemo MCP – Overview”

Persistent, searchable long-term memory for AI agents. Store facts, preferences, and decisions once and recall them across sessions with hybrid natural-language and entity-graph search.

What it does

Section titled “What it does”
  • Store and recall: add memories with categories and tags; search by natural-language query, browse, update, and delete.
  • Entity knowledge graph: relate memories to entities, walk a neighbourhood subgraph, view an entity timeline, and take a point-in-time (as_of) view of what was known at a given moment.
  • Lifecycle: archive and restore, LLM-based consolidation and compression, and JSONL export/import for portability.

Tools

Section titled “Tools”

Eleven single-purpose memory tools (the primary interface), the deprecated memory dispatcher that mirrors them and adds the entity graph, timeline, and point-in-time actions, and config for status, runtime settings, model warmup, and embedding backfill. Every action and parameter is listed in the tools reference.

Deployment

Section titled “Deployment”

mnemo is self-hosted and speaks Streamable HTTP only, on a single /mcp endpoint. Run an instance locally with uvx --from mnemo-mcp mnemo-mcp (loopback, no auth), as an always-on docker compose stack, or on a host of your own behind TLS with token or per-user (multi) auth. Any MCP client that supports HTTP servers connects to it – Claude Code (with or without the plugin), Codex, Gemini CLI, Cursor, Windsurf, and claude.ai web for a remote instance.

All state is one SQLite database under ~/.mnemo/; back it up or move it with any file tool, or use export_memories / import_memories.

Configuration

Section titled “Configuration”

Auth and providers live in the host-owned ~/.mnemo/config.toml. Each task – embedding, reranking, chat (compression, consolidation, importance scoring), and scoring – has its own [models.*] cell with an OpenAI-compatible base_url, api_key, and model, so cloud and local providers can be mixed. With no embedding or rerank cell, mnemo uses local Fastretrieval ONNX models and needs no API key; an unconfigured cell never selects a paid provider. Full walkthrough in setup.

Next steps

Section titled “Next steps”