One binary that stops your coding agent from forgetting everything

What it does

Mnemon is a single Go binary that adds persistent memory to AI agents. It maintains a four-graph knowledge store spanning temporal, entity, semantic, and causal relationships, with intent-aware recall, importance decay, and automatic deduplication. The agent’s host LLM supervises what gets stored, linked, and forgotten, while the binary handles the deterministic work of storage, graph indexing, search, and decay.

The interesting bit

Instead of embedding an LLM inside the memory pipeline like Mem0 or Letta, Mnemon treats your existing subscription as the intelligence layer. The binary exposes three intent-native primitives—remember, link, recall—that map to cognitive vocabulary rather than database jargon, so the model issues judgment calls and the binary executes the storage. This keeps inference costs on the host you already pay for.

Key highlights

• One local binary with no separate API keys; deterministic storage, indexing, and search run locally
• Four-graph knowledge store with intent-aware recall, importance decay, and automatic deduplication
• Supports Claude Code, Codex, Cursor, DeepSeek Harness, OpenClaw, and a dozen other runtimes via native hooks
• Memory compounds over time: the longer it runs, the more continuity it preserves
• Agency Preview adds durable, project-local responsibility admission for Pi agents (macOS/Linux only)

Caveats

• Agency Preview is currently Pi-first and unavailable on Windows until native security boundaries are implemented
• DeepSeek Harness integration requires the separate dsh-mnemon plugin rather than the built-in mnemon setup flow

Verdict

Worth a look if you juggle multiple AI coding agents and want a single, private memory layer that compounds over time. Less compelling if you already have a runtime-native memory you trust, or if you need Agency features on Windows today.