# Mnemon

A single local binary that gives AI agents persistent, cross-session memory as a graph-based store, with intent-aware recall, importance decay and automatic deduplication, supervised by the host LLM rather than an embedded model.

Mnemon stores memory as a four-graph knowledge store and exposes three primitives, `remember`, `link` and `recall`, as an intent-native protocol for how an LLM interacts with a memory database, with structured JSON output rather than raw rows. The binary handles deterministic computation such as storage, graph indexing, search and decay, while the host LLM decides what to remember, how to link it and when to forget it, with no extra inference cost from an embedded model.

It ships one executable with two surfaces: memory operations at the `mnemon` root, and an optional "Agency Preview" at `mnemon agency` that adds durable, project-local responsibility and effect admission to an existing Pi agent, without replacing memory or the underlying agent runtime.

It works with DeepSeek Harness, Claude Code, OpenClaw and other agent runtimes, needs no API keys, and is written in Go.

Website: https://github.com/mnemon-dev/mnemon