mneme

# Mneme [](https://github.com/tr0mb1r/mneme/actions/workflows/ci.yml) [](https://github.com/tr0mb1r/mneme/actions/workflows/cross-build.yml) [](LICENSE) > A standalone, MCP-native memory tool for any LLM or agent. > Single binary. Local-first. Rust. Built to last. ## What it is Mneme is a persistent memory tool for AI agents. It runs as a long-lived process on your machine, exposes its functionality via the Model Context Protocol (MCP), and lets any compatible agent — Claude Desktop, Claude Code, Cursor, Cline, Aider — remember things across sessions. The clearest one-line description: **Mneme remembers things about your work that the agent would otherwise forget.** ## What it isn't - A vector database (it uses one internally, but that's an implementation detail) - A RAG framework - A codebase indexer (modern agents read code with shell tools — that's not Mneme's job) - A web service or SaaS - A library to embed in another application - An LLM ## Setup - **Claude Code** — see [`docs/CLAUDE_CODE_SETUP.md`](docs/CLAUDE_CODE_SETUP.md) for the full guide (recommended path for terminal users). - **Claude Desktop / other MCP hosts** — see [Smoke-testing the MCP server](#smoke-testing-the-mcp-server) below. - **Understanding what mneme actually stores** — see [`docs/MEMORY_LAYERS.md`](docs/MEMORY_LAYERS.md) for a per-layer walkthrough of hot/warm/cold tiers, embedding cadence, and snapshot schedules. - **What isn't built yet** — see [`book/src/roadmap.md`](book/src/roadmap.md). Notably: **on Windows only `mneme run` works** — `mneme daemon`, `mneme client`, and `mneme stop` need Unix domain sockets, so configure your MCP host with `args: ["run"]` there. ## What works today | Layer | Tools | Resource | Storage | |-------|-------|----------|---------| | L0 procedural (always-on) | `pin`, `unpin` | `mneme://procedural` | JSONL on disk, hot-reloaded | | L1 working session | (live state) | `mneme://session/{id}` | `~/.mneme/sessions/<id>.snapshot` | | L3 episodic (recent events) | `recall_recent`, `summarize_session`, `record_event` | `mneme://recent` | redb hot tier + zstd cold quarters | | L4 semantic (long-term facts) | `remember`, `recall`, `update`, `forget` | — | redb + WAL + HNSW vector index | | Auto-context | — | `mneme://context` | All four layers, packed to a token budget | | Diagnostics | `stats`, `list_scopes`, `export`, `switch_scope` | `mneme://stats` | — | The mneme MCP server speaks JSON-RPC against MCP protocol `2025-06-18`, advertises a focused MCP tool and resource surface (see [`book/src/mcp-surface.md`](book/src/mcp-surface.md) for the authoritative inventory), and survives malformed JSON, oversize frames, and EOF cleanly. Real BGE-M3 / MiniLM embeddings via `candle`, HNSW recall via `instant-distance`, atomic snapshots, WAL crash-recovery, schema migration, and `mneme backup` / `mneme restore` round-trips are all in place. v1.1 adds **daemon mode** (`mneme daemon` + `mneme client` bridge — one warm process serving multiple MCP hosts concurrently over a Unix domain socket), the per-agent installer **`mneme init <agent>`** wired for `claude-code`, `claude-desktop`, `cursor`, and `opencode`, and **content size guardrails** with a first-boot upgrade audit. v1.2 adds **encryption at rest** (opt-in, XChaCha20-Poly1305 AEAD over every disk surface; enable with `mneme encrypt`, recover on a fresh machine with `mneme recover`; a 12-word BIP39 recovery phrase is shown once on setup) and **time-range bounds on `recall_recent`** (`since`/`until` parameters, RFC3339 or ULID). v1.3 makes `mneme://context` return all four layers (read `mneme://context?q=<topic>` to get semantic hits), widens the HNSW probe so **filtered `recall` stops underfilling**, adds `tags` / `min_similarity` filters and a `similarity` field on results, **per-project scoping** from the MCP client's workspace roots (`[scopes] derive_from_roots`, off by default), a near-duplicate advisory on `remember`, `resources/templates/list`, and behaviour-hint annotations and titles on every tool — see [`book/src/release-notes-v1_3.md`](book/src/release-notes-v1_3.md). Optional Claude Code lifecycle hooks (`SessionStart`/`PreCompact`/`Stop`) are documented in [`docs/CLAUDE_CODE_SETUP.md`](docs/CLAUDE_CODE_SETUP.md) §7 with ready-to-copy scripts in [`docs/examples/claude-code-hooks/`](docs/examples/claude-code-hooks/). ## Installing Three install paths, each producing the same `mneme` binary on your `$PATH`. See the [installation page](https://tr0mb1r.github.io/mneme/installation.html) in the user docs for the full walkthrough. ### Homebrew (macOS, Linux) — recommended ```sh brew tap tr0mb1r/mneme brew install mneme ``` Pre-built static binary; no Rust toolchain. Apple Silicon, Intel macOS, and aarch64 / x86_64 Linux (musl-static). ### `cargo install` ```sh cargo install mneme-mcp ``` The crate is `mneme-mcp` on crates.io (the bare `mneme` name is held by an unrelated event-sourcing library); the installed binary is `mneme`. Requires Rust stable. ### From source ```sh git clone https://github.com/tr0mb1r/mneme && cd mneme scripts/install.sh # build, install on $PATH, scaffold ~/.mneme scripts/install.sh --minilm # same, but default to MiniLM (~80 MB) instead of BGE-M3 (~1.5 GB) ``` `scripts/install.sh` picks `~/.local/bin` (or `/usr/local/bin` if writable), runs `cargo build --release`, copies the binary, runs `mneme init`, and prints the exact `claude mcp add` line for the next step. Idempotent — safe to re-run when you pull. Pass `--prefix <dir>` to install elsewhere or `--no-init` to skip the data-directory scaffold. ### After install Both Homebrew and `cargo install` skip `mneme init` by design (a formula and `cargo install` should not modify `$HOME`). Run it once manually: ```sh mneme init ``` `mneme init` writes `~/.mneme/config.toml` with all defaults made explicit; edit it before first run if you want to override the embedding model, data directory, or storage budget. The first `mneme run` downloads the embedding model: | Model | Size | Speed | Recall | When to pick | |-------|------|-------|--------|--------------| | `bge-m3` (default) | ~1.5 GB | slower cold start | top-tier, multilingual | You want the best recall and don't mind the disk + first-boot wait. | | `minilm-l6` | ~80 MB | sub-second cold start | good for English | You want fast onboarding, English-only is fine, or you're testing before committing to BGE-M3. | Switching models later re-embeds every stored memory automatically; no manual reindex. Building from source requires Rust stable, pinned via `rust-toolchain.toml`. ## Smoke-testing the MCP server Once built, you can drive `mneme run` with any MCP host. To verify manually with Claude Desktop on macOS: 1. Note the binary's absolute path: `$(pwd)/target/release/mneme` 2. Add it to `~/Library/Application Support/Claude/claude_desktop_config.json`: ```json { "mcpServers": { "mneme": { "command": "/absolute/path/to/target/release/mneme", "args": ["run"] } } } ``` 3. Restart Claude Desktop. The tools panel should list `mneme` with its full tool inventory (see [`book/src/mcp-surface.md`](book/src/mcp-surface.md)) and resource set. The first call may take a few seconds while the embedding model loads. To smoke from the shell without an MCP host: ```sh { printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"shell","version":"0"}}}' printf '%s\n' '{"jsonrpc":"2.0","method":"notifications/initialized"}' printf '%s\n' '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' } | ./target/release/mneme run 2>/dev/null ``` You should see two JSON lines: an `initialize` response advertising the server's capabilities, then a `tools/list` response enumerating every registered tool (see [`book/src/mcp-surface.md`](book/src/mcp-surface.md) for the canonical list). For the comprehensive end-to-end check (every tool, backup/restore round-trip, post-restore recall) run: ```sh scripts/manual_test.sh --stub # offline, ~10s, no model download scripts/manual_test.sh # real MiniLM, exercises the embedder ``` ## Building ```sh cargo build --release ./target/release/mneme --help ``` ## Contributing See [CONTRIBUTING.md](CONTRIBUTING.md). The published docs are the contract: if behaviour and docs disagree, that's a bug in one of them, and the fix belongs in the same commit as the change that caused it. [`book/src/roadmap.md`](book/src/roadmap.md) lists what's deliberately not built yet. ## License Apache-2.0. See [LICENSE](LICENSE).