Agent memory stops relying on hand-written rules: MemSkill turns add, update, delete and skip into learnable, evolving skills
Synopsis
Large language model agents have built memory with a small set of static, hand-designed operations that adapt poorly to varied interaction patterns. MemSkill reframes extraction, consolidation and pruning as learnable, evolvable memory skills: a controller picks a few relevant skills per text span, an LLM executor generates skill-guided memories in one pass, and a designer periodically refines and adds skills from hard cases. On LoCoMo, LongMemEval, HotpotQA and ALFWorld it improves task performance over strong baselines in both conversational QA and embodied interaction, and the skill bank transfers across base models and datasets.
Interpretation
Agent memory construction moves from a fixed hand-designed pipeline to a generation process driven by a reusable, evolving set of memory skills. Earlier memory systems mostly relied on static, hand-designed operation primitives such as add, update, delete and skip, plus heuristic modules deciding what to store, how to revise it and when to prune, and these degrade under diverse interaction patterns and growing histories. On the LoCoMo and LongMemEval conversational benchmarks, MemSkill attains the best LLM-judge scores within both the LLaMA-3.3-70B-Instruct and Qwen3-Next-80B-A3B-Instruct blocks, for example 53.82 on LoCoMo and 60.89 on LongMemEval in the LLaMA block.
The memory skills themselves evolve during training, rather than only learning to use a fixed set. Learning only skill selection, or only allowing existing skills to be refined without adding new ones, leaves uncovered failure modes. In the LoCoMo ablation, replacing the controller with random skill selection lowers the LLM-judge score to 48.43 on LLaMA and 42.84 on Qwen, disabling the designer and fixing the four initial primitives gives 46.50 and 36.15, and refinement-only gives 47.45 and 48.88, all below the full setting's 53.82 and 54.14.
The skill bank builds memory span by span rather than turn by turn during training, cutting repeated LLM calls. Earlier pipelines interleaved handcrafted operations with LLM calls to extract and revise memory incrementally turn by turn, which is costly on long histories. In the LoCoMo cost analysis, span size 512 gives an LLM-judge score of 53.82 with 249K input tokens, 18K output tokens and 215 calls, versus 48.64 with 1288 calls for MemoryOS, 49.71 with 1548 calls for A-MEM and 51.95 with 685 calls for LightMem.
Perspective
The demonstrated scope is evaluation on four benchmarks, LoCoMo, LongMemEval, HotpotQA and ALFWorld, covering conversational long-history memory and embodied interaction tasks, with LLaMA-3.3-70B-Instruct and Qwen3-Next-80B-A3B-Instruct as base models and additional conversational results on Llama-3.1-8B-Instruct. The skill bank is learned and evolved before deployment, and memory is built span by span at runtime with a default span size of 512 tokens.
Performance is mildly sensitive to the number of selected skills K: on HotpotQA across the three context-length settings, K=7 gives the best results, and smaller K may under-use the skill bank in longer contexts. Span size shows a non-monotonic quality-cost relation, since smaller spans may fragment events and larger spans may overlook details. The benefit of skill evolution depends on the stabilized-reward and rollback mechanisms used during training.
