Mnema

ai.zimacPublished by zimac.aiv0.2.0Updated Oct 24, 2026

Local-first, formally verified memory for AI agents: consult/learn/related/list over MCP.

VerifiedSTDIODesktop onlyKnowledge & MemoryAI & ML

Overview

AI-generated overview

Gives AI agents a local, persistent memory store they can consult, add facts to, and browse over MCP.

What you can do
Mnema runs a shared local memory store for AI agents and exposes four MCP tools: consult (search memory before acting), learn (record a durable fact), related (walk graph neighbours), and list (browse recent entries). Recall combines lexical search with graph spreading-activation and optional semantic embeddings, and answers carry a confidence band. The same store can be shared by multiple clients such as Claude Code, Claude Desktop, Cursor, scripts, and CI.
When to use it
Worth adding when you want an assistant to remember durable facts across sessions and clients without sending that memory to a cloud service. Useful for agents, scripts, and CI that should consult prior learnings before acting. Less relevant if you only need per-session context.
Requirements
Runs as a local process (stdio) on macOS or Linux, installed from an .mcpb bundle or the standalone binary. By default it stores data in ~/.mnema and auto-starts a local daemon at Pointing MNEMA_URL at a remote store requires MNEMA_TOKEN as a bearer credential. Optional environment variables: MNEMA_DATA_DIR, MNEMA_URL, MNEMA_TOKEN.
Before you install
By default data stays local in ~/.mnema, but setting MNEMA_URL to a non-loopback address turns the server into a bridge that sends memory to that remote store, so only point it at a store you trust. A remote store requires the MNEMA_TOKEN secret; keep it out of shared configs. The learn tool writes durable facts into the shared store, so anything recorded is visible to every client using that store.

Installation

In SourceWeft

  1. Open Mnema in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

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):

PlatformAsset
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:

sh
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

sh
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:

json
{ "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:

sh
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

EnvDefaultWhat
MNEMA_DATA_DIR~/.mnemaWhere the store lives
MNEMA_URLhttp://127.0.0.1:7077Store the CLI/MCP bridge targets
MNEMA_TOKEN—Bearer credential for a remote store

Links

Mnema is a Zimac project.

Source: README.md at commit 3ad94ce

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.2.0LatestSep 16, 2026

Score

40% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts

    Declare a license in the repository or the registry entry.