Abstract
Long-term memory lets LLM agents reuse past interactions, but raw dialogue histories are verbose and information-sparse. Retrieving broadly improves evidence coverage yet overwhelms downstream reasoning with noise; compressing at write time reduces noise but irreversibly discards details the future query may need. We introduce LazyMem, which sidesteps this dilemma by deferring all memory construction to query time. A lightweight 4B model processes the retrieved candidate pool in overlapping parallel windows, selectively retaining and compressing only query-relevant content. The model is trained through supervised fine-tuning followed by group-based reinforcement learning with a format-gated composite reward that combines a rule-based action signal measuring selection accuracy with an LLM-judged quality signal measuring source faithfulness and query utility. On the LongMemEval benchmark, LazyMem-4B achieves an LLM-judge accuracy of 0.85 with only 213 memory tokens, 68.7$\times$ fewer than retrieval-only, and generalizes to LoCoMo (0.68) without target-domain training, while reducing mean latency over the prior query-time baseline. The 32B variant reaches 0.93, surpassing oracle-context references on aggregation-heavy question types. The code associated with this work is publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2607.22690
