a2328275243/mempalace-evolve

mempalace-evolve is a Python library that gives AI agents persistent memory across sessions. A typical AI conversation ends and all context disappears. This project stores what the agent learns, scores each piece of memory for usefulness, and automatically cleans up stale or redundant entries over time so the stored memory stays relevant rather than growing into an unmanageable pile. The library organizes memory into three types drawn from cognitive science: semantic (general facts and knowledge), episodic (specific events and conversation history), and procedural (error patterns and lessons learned). Semantic and episodic memories are kept separate per project, while procedural memories are shared across all projects automatically. Two retrieval paths are available. The first is vector-based semantic search using ChromaDB, which finds relevant memories even when the query uses different words than the stored text. The second is a knowledge graph that tracks relationships between entities. The digest() function extracts key information from a conversation automatically, and context_for() surfaces the most relevant memories for a new query so they can be injected into an AI prompt before calling the language model. The self-evolution mechanism runs when you call evolve(). It scores candidate memories, promotes high-quality ones to long-term storage, merges duplicates, and removes entries that score poorly or have not been accessed in a long time. Memories that get recalled frequently stay alive, those that are never retrieved fade out gradually. Installation requires Python 3.10 or later. The only mandatory dependency is ChromaDB, and no GPU is needed. The library connects to AI tools like Claude Code and Cursor through the Model Context Protocol (MCP), or it can be called directly from Python code. An optional REST API is available for connecting to non-Python applications. The README is written in Chinese. The license is MIT.