mnemograph

Mnemograph is a persistent, event-sourced knowledge graph designed for AI coding agents. It stores entities, relations, and observations so agents can perform semantic search, tiered context retrieval, and git-style versioning of memory. The server exposes a rich set of MCP tools you can invoke through any MCP client: remember to store knowledge atomically, recall to retrieve context with optional focus on specific entities, and graph maintenance tools to manage entities, relations, and observations. You can also manage sessions, branches, and time-travel-like operations to view or rewind history. The default recall output is prose-form contextual narration, with an optional graph-format JSON when you request format="graph". Common workflows include creating memory via remember, exploring the graph with read_graph or find_similar, and using git-like operations (vcs) to version-control memory state. The tools are organized into core operations, session lifecycle, branching, graph maintenance, time travel, and edge-weight management, providing end-to-end capabilities for building, querying, and evolving a knowledge graph alongside your AI agents.

Prerequisites:

• Python 3.8+ (recommended latest 3.x)
• Internet access to fetch PyPI packages
• Optional: a Claude Code or other MCP client to wire up mnemograph, or use uvx directly as shown

Install and set up:

1. Install the package from PyPI:

pip install mnemograph

2. Initialize memory path (adjust as needed):

mkdir -p ~/.claude/memory

3. Register the MCP client with Claude Code or your MCP client of choice (examples below).

• For Claude Code (global scope):

claude mcp add --scope user mnemograph \
  -e MEMORY_PATH="$HOME/.claude/memory" \
  -- uvx mnemograph

• For other MCP clients, follow the UNIVERSAL_MCP_COMPATIBILITY guidelines and point the client to the mnemograph command exposed by uvx.

4. Run the server locally via uvx (or integrate into your environment as described above):

uvx mnemograph

Optional: If you prefer installing from source, clone and sync as in the repo README, then hook up your MCP client to the local uvx-based command as shown in the setup steps.

Environment variable MEMORy_PATH controls where the memory graph is stored. Ensure your chosen path is writable by the user running the MCP client. Remember that mnemograph uses event sourcing; frequent commits and proper git-based versioning (vcs) help preserve history. When using recall, you can request a graph format to get structured JSON in addition to the default prose. If you encounter large graphs, consider using shallow or medium recall depths to avoid excessive token usage. For multi-project setups, you can set MEMORY_PATH to a project-local directory or a shared global path as needed; the CLI supports a --global flag to switch scope when appropriate.