Mnemo

Autonomous bitemporal memory engine and AST-based knowledge graph for AI agents via Model Context Protocol (MCP). Zero external infrastructure, single-file SQLite storage.

Core Capabilities

• Bitemporal Fact Store: Separates validity timeline (valid_time) from system recording time (ingest_time) using SQLite WAL mode.

• 3-Channel Hybrid Retrieval: Fuses dense vector embeddings, FTS5 full-text search, and recursive CTE graph traversal via Reciprocal Rank Fusion (RRF).

• Ebbinghaus Memory Decay: Four dynamic tiers (Core, Working, Peripheral, Archived) prevent context bloat through piecewise exponential decay.

• AST Knowledge Graph: Parses project syntax trees, tracks cross-file class/function relationships, and reconciles deleted modules automatically.

• Zero Infrastructure: Operates locally inside a single embedded database file without requiring Docker, external vector databases, or cloud accounts.

Related MCP server: NEAT MCP Server

Installation

Install the package directly from PyPI into your environment:

pip install mnemo-agentmcp

Or run commands directly via uvx without global installation:

uvx --from mnemo-agentmcp mnemo --help

MCP Server Configuration

Connect Mnemo to your preferred AI environment (Cursor, Claude Desktop, Antigravity, Windsurf) by updating your client settings.

Cursor / Antigravity (.cursor/mcp.json or mcp_config.json)

{
  "mcpServers": {
    "mnemo": {
      "command": "python",
      "args": ["-m", "mnemo.mcp.server"]
    }
  }
}

Zero-Setup via uvx (Cross-Platform)

{
  "mcpServers": {
    "mnemo": {
      "command": "uvx",
      "args": ["--from", "mnemo-agentmcp", "mnemo", "serve"]
    }
  }
}

Quick Start

Initialize the memory database in your project directory:

mnemo init

Analyze the repository topology and populate the initial knowledge graph:

mnemo scan .

Generate an interactive dark-themed Sankey visualization of your memory tiers and graph connections:

mnemo visualize

Verify your environment dependencies and database health:

mnemo doctor

Available MCP Tools

Agents communicate with Mnemo over stdio JSON-RPC using these tools:

• mnemo_remember: Classify and persist facts using AUDN operations (Add, Update, Delete, Noop).

• mnemo_search: Retrieve facts using hybrid vector, lexical, and graph ranking.

• mnemo_invalidate: Soft-retire facts by ending their validity window without losing history.

• mnemo_reinforce: Boost fact salience upon confirmation or repeated access.

• mnemo_inspect: View full metadata, source trace, and bitemporal intervals for any fact.

• mnemo_get_debt_ledger: Audit active architectural tensions, deprecated items, and unresolved debt.

• mnemo_scan_project: Trigger project-wide AST dependency re-indexing directly from the agent.

Architecture Overview

License

This project is licensed under the MIT License — see the LICENSE file for details.