mnemon-dev/mnemon

What it solves

LLM agents typically suffer from "memory loss" between sessions, where critical decisions and cross-session knowledge vanish or are pushed out of the context window during compaction. Mnemon provides a persistent, cross-session memory layer that allows agents to maintain a long-term knowledge base across different sessions and even different AI frameworks.

How it works

Mnemon operates as a standalone native binary that handles deterministic storage and indexing, while the host LLM acts as the supervisor. Instead of embedding its own LLM, Mnemon uses the agent's existing intelligence to decide what to remember, how to link information, and when to forget. It integrates into agent runtimes (like Claude Code, Cursor, or Pi) via runtime-specific skills, behavioral guides, and lifecycle hooks (Prime, Remind, Nudge, and Compact) that prompt the agent to interact with the memory store.

Who it’s for

Developers and users of AI agents who want their tools to maintain a consistent memory across projects and sessions without needing additional API keys or complex infrastructure.

Highlights

• LLM-Supervised Architecture: Uses the host LLM for judgment calls, eliminating extra inference costs and the need for separate API keys.
• Four-Graph Knowledge Store: Organizes data using temporal, entity, causal, and semantic edges rather than relying solely on vector similarity.
• Intent-Native Protocol: Uses cognitive commands (remember, link, recall) that map to how LLMs think, rather than standard database syntax.
• Broad Framework Support: Integrates with a wide array of tools including Claude Code, Cursor, Codex, Pi, and DeepSeek Harness.
• Automatic Maintenance: Includes built-in deduplication, importance decay, and garbage collection to keep memory clean.