mnemon

<p align="center"> <img src="docs/logo/logo.svg" width="160" height="160" alt="Mnemon Logo" /> </p> # Mnemon **English** | [中文](docs/zh/README.md) **LLM-supervised persistent memory for AI agents.** [](https://go.dev/) [](https://github.com/mnemon-dev/mnemon/actions/workflows/ci.yml) [](https://goreportcard.com/report/github.com/mnemon-dev/mnemon) [](LICENSE) --- LLM agents forget everything between sessions. Context compaction drops critical decisions, cross-session knowledge vanishes, and long conversations push early information out of the window. Mnemon gives your agent persistent, cross-session memory — a four-graph knowledge store with intent-aware recall, importance decay, and automatic deduplication. Single binary, zero API keys, one setup command. > **Experimental beta:** this repository also includes `mnemon-harness`, a > source-built beta for project-local host-agent lifecycle state. It is separate > from the stable `mnemon` CLI, not production-ready, and may make breaking > changes at any time. See [harness/README.md](harness/README.md). > **Claude Max / Pro subscriber?** Mnemon works entirely through your existing subscription — no separate API key required. Your LLM subscription *is* the intelligence layer. Two commands and you're done. ### Why Mnemon? Most memory tools embed their own LLM inside the pipeline. Mnemon takes a different approach: **your host LLM is the supervisor.** The binary handles deterministic computation (storage, graph indexing, search, decay); the LLM makes judgment calls (what to remember, how to link, when to forget). No middleman, no extra inference cost. | Pattern | LLM Role | Representative | |---|---|---| | **LLM-Embedded** | Executor inside the pipeline | Mem0, Letta | | **File Injection** | None — reads file at session start | Claude Code Memory | | **MCP Server** | Tool provider via MCP protocol | claude-mem | | **LLM-Supervised** | External supervisor of a standalone binary | **Mnemon** | Mnemon also addresses a gap in the protocol stack. MCP standardizes how LLMs discover and invoke tools. ODBC/JDBC standardizes how applications access databases. But how LLMs interact with databases using memory semantics — this layer has no protocol. Mnemon's three primitives — `remember`, `link`, `recall` — form an intent-native protocol: command names map to the LLM's cognitive vocabulary (`remember` not INSERT, `recall` not SELECT), and output is structured JSON with signal transparency rather than raw database rows. <p align="center"> <img src="docs/diagrams/llm-supervised-concept.jpg" width="720" alt="LLM-Supervised Architecture — three patterns compared, with Mnemon hooks, protocol boundary, and deterministic memory engine" /> <br /> _{The LLM-Supervised pattern: hooks drive the lifecycle, the host LLM makes judgment calls, the binary handles deterministic computation.} </p> Memory has a **compound interest effect** — the longer it accumulates, the greater its value. LLM engines iterate constantly, skill files cost nearly nothing to write, but memory is a private asset that grows with the user. It is the only component in the agent ecosystem worth deep investment. <p align="center"> <img src="docs/diagrams/10-knowledge-graph.jpg" width="720" alt="Knowledge Graph — 87 insights connected by temporal, entity, semantic, and causal edges" /> <br /> _{A real knowledge graph built by Mnemon — 87 insights, 2150 edges across four graph types.} </p> See [Design & Architecture](docs/DESIGN.md) for details. ## Quick Start ### Install **Homebrew** (macOS / Linux): ```bash brew install mnemon-dev/tap/mnemon ``` **Go install**: ```bash go install github.com/mnemon-dev/mnemon@latest ``` **From source**: ```bash git clone https://github.com/mnemon-dev/mnemon.git && cd mnemon make install ``` **Verify installation**: ```bash mnemon --version ``` ### Claude Code ```bash mnemon setup ``` `mnemon setup` auto-detects Claude Code, then interactively deploys skill, hooks, and behavioral guide. Start a new session — memory just works. ### [Codex](https://github.com/openai/codex) ```bash mnemon setup --target codex --yes ``` One command deploys the mnemon skill, prompt files, and Codex lifecycle hooks (`SessionStart`, `UserPromptSubmit`, `Stop`) in `.codex/hooks.json`. ### [OpenClaw](https://github.com/openclaw/openclaw) ```bash mnemon setup --target openclaw --yes ``` One command deploys skill, hook, plugin, and behavioral guide to `~/.openclaw/`. Restart the OpenClaw gateway to activate. ### [Pi](https://pi.dev) ```bash mnemon setup --target pi --yes ``` One command deploys the mnemon skill, prompt files, and a Pi TypeScript extension to `.pi/`. The extension maps Mnemon's lifecycle reminders onto Pi events (`resources_discover`, `before_agent_start`, `agent_end`, `session_before_compact`). Start a new Pi session or run `/reload` to activate. ### [Hermes Agent](https://github.com/NousResearch/hermes-agent) ```bash mnemon setup --target hermes --yes ``` One command deploys the mnemon skill, prompt files, and Hermes shell hooks to `~/.hermes/`. The integration uses Hermes' native lifecycle hooks: `on_session_start`, `pre_llm_call`, `post_llm_call`, and optional `on_session_finalize`. Hermes may prompt once to approve the installed shell hooks. ### [NanoClaw](https://github.com/qwibitai/nanoclaw) NanoClaw runs agents inside Linux containers. Use the `/add-mnemon` skill to integrate: 1. Install mnemon on the host (see above) 2. In your NanoClaw project, run `/add-mnemon` — Claude Code will modify the Dockerfile, add a container skill, and set up volume mounts 3. Each WhatsApp group gets its own isolated memory store, with optional global shared memory (read-only) The skill is available at `.claude/skills/add-mnemon/` in the NanoClaw repo. ### [Nanobot](https://github.com/HKUDS/nanobot) ```bash mnemon setup --target nanobot --global --yes ``` One command writes a skill file to `~/.nanobot/workspace/skills/mnemon/SKILL.md`. Memory is shared across all Nanobot sessions and projects. Use `--global` (recommended) because Nanobot discovers skills from the global workspace directory. ### Uninstall ```bash mnemon setup --eject ``` ## How it works Once set up, memory operates through a lightweight harness: `SKILL.md` teaches commands, `GUIDELINE.md` teaches judgment, hooks remind the agent at lifecycle boundaries, and the `mnemon` binary executes deterministic memory operations. Supported setup commands automate this, but the harness is installable from markdown alone. ```text Session starts | v Prime -> make skill, guideline, and active store visible | v User prompt arrives | v Remind -> decide whether recall could change this task | v Agent works and calls Mnemon only when useful | v Nudge -> decide whether durable writeback is justified | v Before context compaction | v Compact -> preserve only critical continuity ``` The four hook phases are reminders, not a hard workflow. **Prime** makes the skill, guideline, and active store visible. **Remind** prompts a recall decision. **Nudge** prompts a writeback decision. **Compact** preserves only critical continuity before context compression. You don't run mnemon commands yourself. The agent does when the guideline says memory is useful. ## Features - **Zero user-side operation** — install once; supported runtimes can use hooks, minimal runtimes can use persistent rules - **LLM-supervised** — the host LLM decides what to remember, update, and forget; no embedded LLM, no API keys - **Multi-framework support** — Claude Code and Codex (hooks), OpenClaw (plugins), Pi (extensions), Nanobot (skills), and more - **Markdown-installable harness** — `SKILL.md`, `INSTALL.md`, `GUIDELINE.md`, and four lifecycle reminders - **Four-graph architecture** — temporal, entity, causal, and semantic edges, not just vector similarity - **Intent-native protocol** — three primitives (`remember`, `link`, `recall`) map to the LLM's cognitive vocabulary, not database syntax; structured JSON output with signal transparency - **Intent-aware recall** — graph traversal + optional vector search (RRF fusion), enabled by default for all queries - **Built-in deduplication** — `remember` auto-detects duplicates and conflicts; skips or auto-replaces - **Retention lifecycle** — importance decay, access-count boosting, and garbage collection - **Privacy-safe receipts** — export hashed operation receipts for memory-boundary audits without raw memory contents or queries - **Optional embeddings** — works fully without Ollama; add local [Ollama](https://ollama.ai) for enhanced vector+keyword hybrid search ## Vision All your local agentic AIs — across sessions and frameworks — sharing one pool of live memory. ``` Claude Code ──┐ │ OpenClaw ─────┤ │ Pi ───────────┤ │ Nanobot ──────┤ │ NanoClaw ─────┤ ├──▶ ~/.mnemon ◀── shared memory OpenCode ─────┤ │ Gemini CLI ───┘ ``` The foundation is in place: a single `~/.mnemon` database that any agent can read and write. Claude Code setup automates hook installation; OpenClaw can use plugin hooks; Pi integrates via native skills and TypeScript lifecycle extensions; Nanobot integrates via skill files; NanoClaw integrates via container skills and volume mounts. The same harness can be installed in any LLM CLI that supports skills, rules, system prompts, or event hooks. The longer-term direction is a **memory gateway**: protocol decoupled from storage engine. The current SQLite backend is the first adapter; the protocol surface (`remember / link / recall`) can sit on top of PostgreSQL, Neo4j, or any graph database. Agent-side optimization (when to recall, what to remember) and storage-side optimization (indexing, graph algorithms) evolve independently. See [Future Direction](docs/design/08-decisions.md#82-future-direction) for details. ## FAQ **Do different sessions share memory?** Yes. By default, all sessions use the same `default` store — a decision remembered in one session is available in every future session. **Can I isolate memory per project or agent?** Yes. Use named stores to separate memory: ```bash mnemon store create work # create a new store mnemon store set work # set as default MNEMON_STORE=work mnemon recall "query" # or use env var per-process ``` Different agents/processes can use different stores via the `MNEMON_STORE` environment variable — no global state contention. **Local or global mode?** `mnemon setup` defaults to **local** (project-scoped `.claude/`), recommended for most users. **Global** (`mnemon setup --global`, installed to `~/.claude/`) activates mnemon across all projects — convenient if you want other frameworks (e.g., OpenClaw) to share memory by forwarding requests through Claude Code CLI, but may add maintenance overhead. **How do I customize the behavior?** Edit the generated guideline (`~/.mnemon/prompt/guide.md` in current setup flows) or use the installable [memory loop GUIDE](harness/loops/memory/GUIDE.md) as the source. The skill file should stay focused on command syntax. **What is sub-agent delegation?** Sub-agent delegation is optional. When a runtime supports it, the main agent can decide *what* to remember and ask a cheaper or isolated worker to execute `mnemon remember`. It is a useful execution strategy, not a required part of the Mnemon architecture. ## Configuration | Environment Variable | Default | Description | |---|---|---| | `MNEMON_DATA_DIR` | `~/.mnemon` | Base data directory | | `MNEMON_STORE` | *(active file or `default`)* | Named memory store for data isolation | **Ollama-specific** (only relevant if using embeddings): | Environment Variable | Default | Description | |---|---|---| | `MNEMON_EMBED_ENDPOINT` | `http://localhost:11434` | Ollama API endpoint | | `MNEMON_EMBED_MODEL` | `nomic-embed-text` | Embedding model name | ## Development ```bash make build # build binary make install # build + install to $GOBIN make test # run E2E test suite mnemon setup # interactive setup mnemon setup --eject # remove all integrations make help # show all targets ``` **Dependencies**: Go 1.24+, `modernc.org/sqlite`, `spf13/cobra`, `google/uuid` See [Development and Deployment](docs/DEPLOYMENT.md) for Docker, Compose, Ollama embedding, and release setup. ## Documentation - [Mnemon Harness Beta](harness/README.md) — experimental host-agent lifecycle state - [Memory Loop Harness](harness/loops/memory/README.md) — installable memory loop assets - [Skill Loop Harness](harness/loops/skill/README.md) — installable skill loop assets - [Design & Architecture](docs/DESIGN.md) — current engine architecture, algorithms, integration design - [Usage & Reference](docs/USAGE.md) — CLI commands, embedding support, architecture overview - [Memory Import Guide](docs/IMPORT.md) — schema and LLM prompt for importing historical chats - [Architecture Diagrams](docs/diagrams/) — system architecture, pipelines, lifecycle management ## Star History <a href="https://www.star-history.com/?repos=mnemon-dev%2Fmnemon&type=date&legend=top-left"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=mnemon-dev/mnemon&type=date&theme=dark&legend=top-left" /> <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=mnemon-dev/mnemon&type=date&legend=top-left" /> <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=mnemon-dev/mnemon&type=date&legend=top-left" /> </picture> </a> ## References Mnemon combines the paradigm of one paper with the methodology of another, grounded in the structural insight that graph memory is isomorphic to LLM attention. See [Theoretical Foundations](docs/DESIGN.md#25-theoretical-foundations) for details. - **RLM** — Zhang, Kraska & Khattab. [Recursive Language Models](https://arxiv.org/abs/2512.24601). 2025. Establishes the paradigm: LLMs are more effective as orchestrators of external environments than as direct data processors. - **MAGMA** — Zou et al. [A Multi-Graph based Agentic Memory Architecture](https://arxiv.org/abs/2601.03236). 2025. Provides the methodology: four-graph model (temporal, entity, causal, semantic) with intent-adaptive retrieval. - **Graph-LLM Structural Insight** — Joshi & Zhu. [Building Powerful GNNs from Transformers](https://arxiv.org/abs/2506.22084). 2025; and the Graph-based Agent Memory survey (Chang Yang et al., 2026). Confirms that LLM attention is computationally equivalent to GNN operations — graph memory is a structural match, not an engineering convenience. ## License Copyright 2026 Grivn and Mnemon contributors. [Apache-2.0](LICENSE)