mnemon-dev/mnemon

What it solves

Mnemon provides AI agents with persistent, cross-session memory, preventing them from forgetting critical decisions and knowledge between sessions. It solves the problem of context window limits and context compaction, where early information is often lost as conversations grow.

How it works

Unlike tools that embed an LLM inside their memory pipeline, Mnemon uses an "LLM-supervised" approach. A standalone binary handles the deterministic storage, graph indexing, and search, while the host LLM acts as the supervisor, making judgment calls on what to remember, how to link information, and when to forget.

It utilizes a four-graph architecture (temporal, entity, causal, and semantic edges) and an intent-native protocol with three primitives: remember, link, and recall. The system is integrated into agent runtimes via a harness consisting of skill files, guidelines, 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 agent frameworks such as Claude Code, Cursor, Codex, TRAE, Kimi Code, and others who want their agents to maintain a consistent, shared memory pool across different projects and sessions.

Highlights

• LLM-Supervised Architecture: The host LLM makes the decisions; the binary handles the computation, eliminating the need for separate API keys or embedded LLMs.
• Four-Graph Knowledge Store: Uses temporal, entity, causal, and semantic edges rather than relying solely on vector similarity.
• Broad Framework Support: One-command setup for a wide array of agentic IDEs and CLIs.
• Intent-Native Protocol: Commands map to the LLM's cognitive vocabulary (remember, recall) rather than database syntax.
• Automatic Maintenance: Includes built-in deduplication, importance decay, and garbage collection.
• Optional Embeddings: Supports local Ollama integration for hybrid vector+keyword search.