mnemon-dev/mnemon
LLM-supervised persistent memory for AI agents — graph-based recall, cross-session knowledge, single binary
Mnemon is LLM-supervised persistent memory for AI agents — a local binary with zero API keys that gives your agent a four-graph knowledge store that survives session boundaries. Most memory tools embed their own LLM; Mnemon takes a different approach: your host LLM is the supervisor, Mnemon handles deterministic computation (storage, graph indexing, search, decay). Three intent-native primitives (remember, link, recall) map to the LLM's cognitive vocabulary — the protocol layer that neither MCP nor ODBC addresses. A real knowledge graph built by Mnemon has 87+ insights and 2,150+ edges. Memory accumulates compound interest: the longer it runs, the more valuable it becomes. Agency Preview (separate surface) adds durable project-local responsibility and effect admission to Pi agents. Claude Max/Pro subscribers use Mnemon entirely through their existing subscription — no extra inference cost.
Install
brew install --cask mnemon-dev/tap/mnemonmacOS: use Homebrew Cask above. Go install (any platform): go install github.com/mnemon-dev/mnemon@latest. From source: git clone https://github.com/mnemon-dev/mnemon.git && cd mnemon && make install. Verify: mnemon --version. Then configure for your agent: mnemon setup (auto-detects Claude Code) or mnemon setup --target codex|cursor|trae|qoder --yes. Windows: core Memory commands only; Agency requires macOS/Linux.
Compatibility
Works with DSH (DeepSeek Harness), Claude Code, Codex, OpenClaw, Cursor, TRAE, Qoder, and any agent runtime that can call shell tools. No API keys required — uses your existing LLM subscription. macOS/Linux for full Agency support; Windows for Memory only.
Details
- Repo: mnemon-dev/mnemon
- Category: Coding & Development
- Stars: 494
- Version: see github.com/mnemon-dev/mnemon/releases for latest. Agency Preview is a separate surface at mnemon agency...
- Last push: 2026-09-17
- First seen: 2026-02-20
Recent updates
Agency Preview available on macOS/Linux. Windows: Memory commands only. Four graph types: insights graph, edges graph, importance decay graph, deduplication graph. Three primitives: remember, link, recall — structured JSON output with signal transparency.
FAQ
- How do I set up Mnemon with DSH?
- Install Mnemon (brew install --cask mnemon-dev/tap/mnemon or go install github.com/mnemon-dev/mnemon@latest), then run mnemon setup with your DSH profile. Mnemon works with any agent runtime that can call shell tools — including DSH.
- Does Mnemon need its own API key or LLM subscription?
- No. Mnemon uses your existing LLM subscription (Claude Max/Pro, etc.). The binary handles deterministic computation; your host LLM makes judgment calls about what to remember, link, and forget. No middleman, no extra inference cost.
- What is the LLM-Supervised pattern?
- Unlike LLM-Embedded tools (Mem0, Letta) or File Injection (Claude Code Memory), Mnemon is LLM-Supervised: the binary handles storage and graph computation deterministically, while your host LLM supervises — deciding what to remember and how to link — without being embedded inside the memory pipeline.
Alternatives
csyangwen/dsh-memory-evolve · Jungod1121/dsh-anchored-standard · bowenliang123/dsh-context