LightMem

LightMem is a Python memory layer for language model agents that targets the cost of building and maintaining memory, not only of answering from it. Conversation turns are compressed and split into topic segments before a model extracts memory entries, which are stored and later retrieved to answer questions about earlier sessions.

The pipeline is modular. A pre-compressor (LLMLingua-2 or entropy-based) shortens incoming messages, a topic segmenter groups them, a memory manager model extracts entries, and an embedder plus retriever (Qdrant, FAISS, or BM25) serves them back. Updates can run offline in a batch after a session instead of on every write, which reduces the model calls spent during memory construction. Each stage is selected through a configuration object.

The memory manager runs on cloud APIs such as OpenAI and DeepSeek or on local models through Ollama, vLLM, or Transformers. An MCP server exposes the memory operations as tools, and tutorial notebooks cover a travel-planning agent, a code assistant, and LongMemEval evaluation.

The repository also hosts related memory methods from the same group: StructMem (hierarchical event-level memory), FluxMem (memory as an evolving heterogeneous graph), and EM²Mem (event-centric multimodal memory for long video). A baseline evaluation toolkit benchmarks memory layers such as Mem0, A-MEM, LangMem, and MemoryOS on LoCoMo and LongMemEval, reporting accuracy, token use, model calls, and runtime.