ReMe

ReMe (Remember Me, Refine Me) is an open-source memory management framework for AI agents, licensed under Apache 2.0. It tackles two core problems of agent memory: limited context windows (where early information is truncated in long conversations) and stateless sessions (where new sessions cannot inherit history). ReMe provides both a file-based memory system (ReMeLight) and a vector-based memory system, achieving state-of-the-art results on the LoCoMo and HaluMem benchmarks.

• File-based memory (ReMeLight): Treats memory as readable, editable Markdown files stored in a structured directory — no opaque databases, easy to inspect and migrate.
• Context management: Automatically checks context size, compacts conversation history into structured summaries, and handles long tool outputs to prevent context overflow.
• Vector-based memory system: Manages personal, procedural, and tool memories using vector stores (local, Chroma, Qdrant, Elasticsearch, or OBVec) with semantic retrieval.
• Hybrid memory search: Combines vector search (weight 0.7) and BM25 keyword matching (weight 0.3) for balanced semantic and exact-match retrieval.
• Pre-reasoning hook: A unified entry point that wires all memory components together — compact tool results, check context, generate summaries, and persist memory asynchronously before each agent reasoning step.
• Persistent long-term memory: Uses a ReAct + file tools pattern so the AI decides what to write and where, persisting important information to daily Markdown journals.
• In-session memory (ReMeInMemoryMemory): Token-aware memory management with raw conversation persistence to JSONL files, supporting session state serialization and deserialization.
• Benchmark performance: Achieves 86.23% overall on LoCoMo and 88.78% QA accuracy on HaluMem, outperforming Mem0, MemOS, Zep, and other baselines.
• Easy installation: Install via pip install reme-ai or from source; configure LLM and embedding backends through environment variables.
• Framework integrations: Integrates with AgentScope, QwenPaw (CoPaw), and supports OpenAI-compatible LLM APIs and multiple vector store backends.