agentmemory - A Persistent Memory System for AI Coding Agents

A persistent memory system that solves the problem of AI coding agents forgetting previous context once a session ends.

It automatically captures and compresses tool usage in the background, then reinjects the necessary context at the start of the next session. As a result, it reduces the need to re-explain your stack and architecture every time, and also eases the CLAUDE.md 200-line limit and stale documentation problems.

• Compatible with MCP/REST-based agents such as Claude Code, Cursor, Codex CLI, Gemini CLI.
• Reported R@5 95.2% on LongMemEval-S.
• Claims to reduce 22K+ tokens to about 1,900 tokens for 240 observations.
• Retrieval combines BM25 + Vector + Graph via RRF.
• Supports @node-rs/jieba and tiny-segmenter for Korean/CJK processing.
• Uses a 4-tier memory structure of Working / Episodic / Semantic / Procedural.
• Provides 51 MCP tools including memory_recall, memory_save, memory_smart_search.
• Auto-detects embeddings ranging from local (all-MiniLM-L6-v2) to Gemini, OpenAI, Voyage, Cohere, OpenRouter.

It also offers a real-time viewer and session replay, and Claude Code JSONL transcripts can be imported via import-jsonl. However, comments mentioned that the current version's post-session hook is not yet connected, and server latency of around 3 seconds.