Mnema

Local-first, formally verified memory for AI agents — one shared store, every client, over MCP.

Your agents rent their intelligence from a lab; their memory should be yours. Mnema is a single, durable store of what your agents have learned — consult it before you act, learn a durable fact, walk the related graph, list what's recent. Data lives in ~/.mnema on your machine and nothing leaves it unless you point it somewhere.

• Verified engine. The embedded engram database is formally specified and machine-checked (property catalog discharged by 250+ tests, an exhaustive CRDT conformance model checker, Kani proofs, and a TLA⁺ model). Every released binary passed that gate on its exact commit.
• Local-first. An open, append-only segment format in ~/.mnema. No cloud, no account.
• One store, every client. Claude Code, Claude Desktop, Cursor, your scripts, and your CI all talk to the same memory. Recall is lexical + graph spreading-activation (plus optional semantic embeddings), with an honest confidence band on every answer.

This repository hosts the public release bundles (.mcpb) and prebuilt binaries. The product, docs, and downloads live at mnema.zimac.ai.

Install

From the MCP Registry (recommended)

Mnema is published to the official MCP Registry as ai.zimac/mnema. Any registry-aware client can install it by name; it pulls the platform .mcpb bundle from this repo's releases and verifies its fileSha256.

.mcpb bundle (manual)

Grab the bundle for your platform from the latest release and open it with an MCP-bundle-aware client (e.g. Claude Desktop → Extensions):

Platform	Asset
macOS (universal)	mnema-<version>-macos-universal.mcpb
Linux (x86_64)	mnema-<version>-linux-x86_64.mcpb

A .mcpb is a ZIP of a manifest.json, the mnema binary, and this README. Each is dry-run verified (the bundled binary completes an MCP initialize handshake) before release.

Headless binary (curl | sh)

For agents, scripts, and CI — installs the standalone mnema binary:

curl -fsSL https://mnema.zimac.ai/install.sh | sh

(macOS universal + Linux x86_64 prebuilt; anything else builds from source. The macOS GUI app is a separate download at mnema.zimac.ai.)

Give your agents memory

claude mcp add mnema -- mnema mcp

That's the whole setup: mnema mcp auto-starts the shared local store if it isn't running, and the session gets four tools — consult (search before you act), learn (record a durable fact), related (graph neighbours), list (browse). Any MCP client works; the generic config is:

{ "mcpServers": { "mnema": { "command": "mnema", "args": ["mcp"] } } }

Point it at a shared / team store

The same binary bridges to a remote Mnema (a token-protected mnema serve or a hosted learnings pod) instead of the local one:

MNEMA_URL=https://memory.example.com MNEMA_TOKEN=<secret> mnema mcp

With a non-loopback URL nothing is auto-started — it's a pure authenticated bridge.

Configuration

Env	Default	What
MNEMA_DATA_DIR	~/.mnema	Where the store lives
MNEMA_URL	http://127.0.0.1:7077	Store the CLI/MCP bridge targets
MNEMA_TOKEN	—	Bearer credential for a remote store

Links

• Product & downloads: https://mnema.zimac.ai
• Registry entry: ai.zimac/mnema on https://registry.modelcontextprotocol.io
• Releases (.mcpb + binaries): https://github.com/Zimac-AI/mnema/releases

---

Mnema is a Zimac project.