mnemon-dev/mnemon

Mnemon – Persistent, LLM‑supervised memory for AI agents

What it is – Mnemon is a small, cross‑platform command‑line tool (a single native binary) that gives large‑language‑model (LLM) agents a durable, searchable knowledge store. The store lives locally in ~/.mnemon and is built as a four‑graph (temporal, entity, causal, semantic) SQLite database.

How it works – The host LLM (Claude, OpenAI, etc.) remains the supervisor: it decides what to remember, how to link facts, and when to forget. Mnemon itself performs only deterministic operations – storing, indexing, graph traversal, decay, deduplication, and optional vector‑search using embeddings. The LLM is never embedded; no extra API keys are needed beyond the user’s existing LLM subscription.

Key features

• Zero‑runtime friction – Install once (via npm, Homebrew, or go install). Supported runtimes automatically get “hooks” or “skills” that remind the LLM to recall or write back memory.
• Multi‑framework integration – Ready‑made setup commands for Claude Code, Codex, Cursor, ZCode, TRAE, Qoder, CodeBuddy, Kimi Code, OpenCode, Hermes Agent, Pi, NanoClaw, Nanobot, DeepSeek Harness, and more.
• Four‑graph architecture – Stores facts as nodes with four edge types, enabling richer reasoning than plain vector similarity.
• Intent‑native protocol – Three primitives (remember, link, recall) map directly to the LLM’s vocabulary and return structured JSON, not raw SQL rows.
• Importance decay & boosting – Memory items fade unless accessed, with automatic garbage collection.
• Privacy‑safe receipts – Export hashed operation logs for audit without exposing raw content.
• Optional embeddings – Works without embeddings; can plug in local Ollama or any OpenAI‑compatible server for hybrid vector‑keyword search.

Typical workflow

1. Setup – mnemon setup --target <runtime> installs the appropriate skill/guide and lifecycle hooks for the chosen LLM tool.
2. Prime – When a session starts, the hook makes the Mnemon store visible to the agent.
3. Recall reminder – Before the LLM processes a user prompt, Mnemon can suggest a recall query.
4. Agent work – The LLM calls mnemon recall … if it decides the retrieved facts are useful.
5. Nudge – After the LLM finishes, Mnemon may prompt the agent to write back new insights (remember).
6. Compact – Just before context compaction, Mnemon helps prune irrelevant memory.

Why it matters – Standard LLM agents lose context across runs; long conversations push early information out of the prompt window. Mnemon provides a persistent, graph‑rich memory that grows with use, letting agents retain decisions, facts, and causal links over months without extra inference cost.

Getting started

npm install -g @mnemon-dev/mnemon   # installs the binary
mnemon setup --target claude        # example for Claude Code
mnemon store create work           # optional named store
MNEMON_STORE=work mnemon recall "project deadline"   # query

Vision – The authors aim to evolve Mnemon into a protocol‑agnostic “memory gateway” that could sit on top of any graph database (PostgreSQL, Neo4j, etc.), decoupling the LLM‑supervised API from the storage engine.

---

TL;DR – Mnemon is a lightweight, LLM‑supervised local memory engine that plugs into many existing AI coding assistants, giving them a durable, graph‑based knowledge base without requiring extra LLM calls or API keys.