Researchers have built a long-term memory system for AI agents that skips the expensive LLM calls most rivals rely on just to remember what happened.
The system, called MemFit and described in a paper posted to arXiv, saves every conversation turn verbatim into an append-only store instead of summarizing or compressing it first. That insertion step is LLM-free and near-instant. Retrieval is handled the same way: a multi-path search that blends keyword matching, semantic search, and cross-encoder reranking, working across both plain text and caption-augmented multimodal content. On three benchmarks, LoCoMo, MemGallery, and LongMemEval-S, the authors report state-of-the-art results while cutting the time and cost of building memory several-fold.
Most agentic-memory systems treat remembering as a reasoning task, spending an LLM call every time a conversation needs filing, which adds up fast for any agent logging thousands of turns a day. MemFit's bet is that storage and retrieval are mostly mechanical problems, and that saving the AI reasoning for actual reasoning is where the real savings are.
The benchmark wins are worth noting, but the paper is grading its own homework. Whether verbatim storage holds up against the messier, longer-running conversations real products generate is the open question.