MemFit paper proposes LLM-free long-term agentic memory
Tags AI / ML
Researchers Mitchell Piehl and Muchao Ye posted MemFit, a long-term memory system for conversational agents that stores each turn verbatim in an append-only store with near-instantaneous, LLM-free insertion, indexing turns with segment summaries rather than replacing them. MemFit uses an LLM-free multi-path retrieval strategy combining lexical and semantic signals with cross-encoder reranking over caption-augmented episodes in both textual and multimodal settings. On the LoCoMo, MemGallery and LongMemEval-S benchmarks, the authors report state-of-the-art performance while reducing memory construction time and cost several-fold. The paper was submitted to arXiv on October 1, 2026.
Technical significance
Removing LLM calls from memory writes attacks the cost and latency bottleneck that makes long-term agent memory expensive to run at scale. If the reported several-fold reduction holds, persistent memory could become practical for high-volume conversational agents rather than a premium feature.