Mnemon provides persistent, cross-session memory for AI agents, addressing the common issue of LLMs forgetting information between interactions. It offers a unified, four-graph knowledge store with intent-aware recall, benefiting developers and users of CLI agents like Claude Code, OpenClaw, and Nanobot by enabling continuous learning and context retention across sessions.

How It Works

Mnemon employs an LLM-supervised approach, distinguishing itself from LLM-embedded solutions. A standalone binary handles deterministic operations such as storage, graph indexing, search, and decay. The host LLM acts as the supervisor, making judgment calls on what to remember, how to link information, and when to forget, thereby avoiding extra inference costs and leveraging existing LLM subscriptions. The system utilizes a four-graph architecture (temporal, entity, causal, semantic) and an intent-native protocol with primitives like remember, link, and recall, which map directly to an LLM's cognitive vocabulary, outputting structured JSON.

Quick Start & Requirements

Installation is straightforward via Homebrew (brew install mnemon-dev/tap/mnemon), Go (go install github.com/mnemon-dev/mnemon@latest), or from source (git clone https://github.com/mnemon-dev/mnemon.git && cd mnemon && make install). Verify installation with mnemon --version. Setup for specific agents like Claude Code (mnemon setup), Codex (mnemon setup --target codex --yes), and OpenClaw (mnemon setup --target openclaw --yes) is automated with single commands. Requires Go 1.21+.

Highlighted Details

• LLM-Supervised Operation: Host LLM dictates memory actions; no embedded LLM or separate API keys required.
• Multi-Framework Support: Integrates seamlessly with Claude Code (hooks), OpenClaw (plugins), Nanobot (skills), NanoClaw (container skills), and others via a Markdown-installable harness.
• Four-Graph Architecture: Employs temporal, entity, causal, and semantic edges for comprehensive knowledge representation.
• Intent-Native Protocol: Uses remember, link, recall primitives aligned with LLM cognitive vocabulary.
• Intent-Aware Recall: Combines graph traversal with optional vector search (RRF fusion).
• Unified Memory: A single ~/.mnemon database allows all local agentic AIs to share memory across sessions and frameworks.
• Optional Embeddings: Fully functional without Ollama; integrates with local Ollama for enhanced hybrid search.

Maintenance & Community

No specific details regarding maintainers, sponsorships, or community channels (e.g., Discord, Slack) were found in the provided README snippet.

Licensing & Compatibility

The project is released under the MIT license, which is permissive and generally compatible with commercial use and closed-source applications.

Limitations & Caveats

The README does not explicitly state limitations such as alpha status or known bugs. The current SQLite backend is noted as the first adapter, suggesting potential future evolution of the storage engine.