skillfed

A learned memory-skill bank beats fixed insert/update/delete baselines by 5-7 points, on 6x fewer LLM calls

Notes on MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (arXiv:2602.02474) — Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, Wenya Wang · February 2026

Note published · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

MemSkill treats the four hard-coded memory operations — insert, update, delete, skip — that most LLM agent memory systems run on as a starting point, not the final word. It organizes them into a shared skill bank, a set of structured, reusable routines for extracting, consolidating, and pruning information from interaction history. Three learned components run that bank: a controller, trained with reinforcement learning (PPO) and Gumbel-Top-K sampling, picks a small set of relevant skills for each turn without replacement; an executor applies whichever skills got picked to write memory in a single pass over a span of history, rather than turn by turn; and a designer periodically reviews the hard cases — where selected skills produced wrong or incomplete memories — clusters similar failures, and proposes refinements or entirely new skills. That closes the loop: the selection policy and the skill set itself both keep improving over time.

MemSkill beats the strongest baselines tested — MemoryOS, Mem0, A-MEM, and Chain-of-Note — across LoCoMo and LongMemEval (conversational memory QA), HotpotQA (used to test distribution shift across document-context lengths), and ALFWorld (embodied, multi-step tasks). On LoCoMo with LLaMA 3.3 70B it scores 44.21 F1 / 53.82 LLM-judge against 41.39 / 48.64 for the best baseline; on ALFWorld it reaches 77.14% / 83.58% success on seen/unseen tasks against 74.29% / 81.34%. Swap in Qwen3-Next 80B with no retraining and the gains hold: LoCoMo LLM-judge 54.14 vs. 47.37, ALFWorld average success 81.29% vs. 74.02%. An ablation shows why: on the harder Qwen transfer setting, dropping the designer alone drags LoCoMo LLM-judge from 54.14 down to 36.15 — a steeper fall than dropping the controller, so both learned pieces are earning their keep. It's cheaper, too: on LoCoMo, MemSkill runs roughly 249K input tokens and 215 LLM calls, against about 1M tokens / 1,288 calls for MemoryOS and 2.85M tokens / 1,548 calls for A-MEM.

Key numbers

LoCoMo score vs. best baseline (LLaMA 3.3 70B)44.21 F1 / 53.82 L-J vs. 41.39 / 48.64
ALFWorld success, seen / unseen (LLaMA 3.3 70B)77.14% / 83.58% vs. 74.29% / 81.34%
LoCoMo L-J w/o designer, Qwen transfer (ablation)54.14 → 36.15
LoCoMo cost: input tokens / LLM calls249K / 215 vs. up to 2.85M / 1,548 for baselines
LongMemEval L-J, transfer setting (no training)60.89 vs. 56.44 best baseline

Skills related to this research

Google Docs This skill connects your agent to Google Docs through the Docs v1 REST API, letting you read document content, create new docs, and insert or replace text. OAuth permissions default to read-only; write access requires explicit user opt-in at install.★ 13 postfx-composer postfx-composer streamlines the setup and configuration of EffectComposer for sophisticated multi-pass post-processing workflows. This skill guides you through building layered visual effects pipelines, managing render targets, and compositing complex graphics transformations. Perfect for developers working with advanced rendering techniques and real-time graphics optimization.★ 8 axiom-audit-swiftdata axiom-audit-swiftdata equips AI coding assistants with specialized knowledge to identify problematic patterns in SwiftData model definitions. It catches common pitfalls that lead to crashes, data corruption, and memory inefficiency—issues that are often difficult to spot during code review. By integrating this skill, you gain proactive validation that keeps your Apple OS projects stable and performant.★ 1,095 seedance-recipes Seedance-recipes provides production-ready recipe patterns for video content across genres: product, lifestyle, drama, music video, landscape, commercial, animation, and more. Each recipe preserves core creative constraints while inviting customization of subject, camera, lighting, and sound. Use recipes as proven starting shapes, not rigid templates.★ 5,445
Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent Systems Large Language Model (LLM)-based automatic Multi-Agent Systems (MAS) generation has become a crucial frontier for tackling complex tasks. However, existing… Lin et al. · similarity 0.88

Related notes

References

  1. Zhang, H., Long, Q., Bao, J., Feng, T., Zhang, W., Yue, H., & Wang, W. (2026). MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents. arXiv:2602.02474.
  2. Mem0 — production-oriented long-term memory system for LLM agents, used as a baseline in this paper.
  3. MemoryOS — OS-inspired memory management framework for AI agents, used as a baseline in this paper.
  4. A-MEM — agentic memory framework for LLM agents, used as a baseline in this paper.
  5. LoCoMo — long-term conversational memory benchmark used for evaluation in this paper.