Awesome AI Agents › AI Agent Development

mnemon-dev/mnemon

⭐ 574 Go repository created 2026-02-20

Mnemon supplies persistent memory to LLM agents that otherwise forget everything between sessions, lose decisions to context compaction and push early information out of the window in long conversations. Its distinguishing design choice is where the model sits. Tools such as Mem0 or Letta embed an LLM inside the memory pipeline; Mnemon instead treats the host LLM as the supervisor, so the agent already running in the terminal makes the judgement calls about what is worth storing, how new material links to old, and when something should fade, while the binary performs only deterministic work: storage, graph indexing, search, importance decay and deduplication. That removes both a second inference bill and the need for a separate API key, and it lets a Claude Max or Pro subscription act as the intelligence layer. The store is a knowledge graph with four edge types, temporal, entity, semantic and causal, and recall is intent-aware rather than a plain similarity query. The command surface is deliberately small and shaped like the model's own vocabulary, with remember, link and recall in place of INSERT and SELECT, and structured JSON output that exposes the signals behind a result. The authors frame this as a missing protocol layer: MCP standardises tool invocation and ODBC standardises database access, but memory semantics between an LLM and a store had no equivalent. Mnemon is written in Go 1.24, backed by SQLite, installed through Homebrew, go install or source, and integrates with Claude Code, OpenClaw and other CLI agents through hooks and MCP. A separate preview surface, mnemon agency, adds project-local responsibility and effect admission and is unavailable on Windows. Documentation is in English and Chinese, and the licence is Apache 2.0.

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

agent-memoryllm-memorypersistent-memoryknowledge-graphmcpclaude-codegolangsqliteragcli