agentscope-ai/ReMe

Provides dual file-based and vector-based memory architectures that compress long conversations into persistent summaries while enabling hybrid semantic search (vectors + BM25). Automatically manages context windows through a Compactor component, persists agent knowledge across sessions as human-readable Markdown files, and includes a MemorySearch tool for retrieving relevant historical context. Integrates with LLM/embedding APIs and includes a file-watcher system that asynchronously summarizes conversations and caches embeddings.

2,185 stars. Actively maintained with 52 commits in the last 30 days.