Open-source memory server agentmemory released for coding agents

agentmemory is an open-source memory layer that helps coding agents maintain task context even after a session ends. It is designed to solve the token cost and context window limitations that arise from injecting the full context every time.

Key Features and How It Works:

• Automatic Data Collection: Through the agent's hooks, it automatically records tool call results, file access patterns, and failure context, while filtering out sensitive information such as API keys.
• 4-Tier Memory Structure: Inspired by the structure of the human brain, it manages information across Working, Episodic, Semantic, and Procedural memory tiers.
• Hybrid Search: It uses a hybrid approach combining BM25 keyword matching, vector search, and graph traversal to accurately retrieve needed information.
• Broad Tool Support: It can connect to various agents that support MCP or REST APIs, including Claude Code, Cursor, and GitHub Copilot CLI.

Performance and Technical Details:

• According to its own benchmark results, the hybrid search achieved high retrieval accuracy (R@5 95.2%).
• For Korean-language users, it supports separate tokenization settings via jieba and tiny-segmenter.