Source: arXiv — 2026-07-18

Summary

Researchers (including authors from MemTensor, USTC, and Hong Kong Polytechnic University) introduce MSCE (Memory-Skill Co-Evolution), a training-free framework that stops treating an LLM agent's memory as passive text dumped back into its context window and instead converts proven experience into callable "skills." Each skill retains the evidence trace it came from, the conditions under which it applies, verification rules, and a reliability estimate, and MSCE promotes a memory into a skill only once accumulated evidence shows it reliably helps. The mechanism for deciding what counts as "proven" is reflection-weighted value backfilling, which spreads a sparse end-of-task reward (success/failure) back across the dense chain of self-reflections an agent generates mid-task. MSCE posts the best or tied-best Pass@1 across all five EvoAgentBench domains and beats the strongest skill-augmented baseline, SkillFlow-Evolve, by 2.01 points on LoCoMo's judge score and 1.18 on F1.

Key Takeaways

• The core distinction: most agent memory systems retrieve a past trace and paste it into the prompt as extra context ("here's what happened last time"). MSCE instead promotes that trace into a skill object — evidence links, applicability boundaries, decision guidance, and a reliability score — that the agent can call directly.
• Reflection-weighted value backfilling is the credit-assignment trick: a task only produces one final reward signal (did it succeed?), but the agent made many intermediate decisions along the way. This method propagates that single reward backward through the agent's own self-reflections to estimate which specific steps actually deserved credit or blame.
• A memory only "graduates" into a reusable skill once its estimated value, backed by real evidence, clears a threshold — this is meant to stop low-quality or lucky one-off traces from calcifying into bad habits the agent keeps repeating.
• Benchmarked against both memory-only and skill-only baselines on EvoAgentBench (five domains) and LoCoMo (long-context conversational memory), MSCE wins or ties for best Pass@1 everywhere and outperforms the strongest skill-augmented competitor, SkillFlow-Evolve, on LoCoMo's judge score (+2.01) and F1 (+1.18).
• It's training-free — no fine-tuning or weight updates — meaning the governance logic sits entirely in how memory is scored, filtered, and promoted, not in the model itself.

Reel Script

Hook (~18s, ~40 words): Most "agent memory" today is just a transcript getting pasted back into the prompt — the agent re-reads what happened and re-decides from scratch every time. A new framework called MSCE instead turns proven experience into something the agent can just call, like a saved function.

Core Concept (~70s, ~155 words): Here's the actual problem. An agent finishes a long task and gets one signal back — success or fail. But that task involved dozens of intermediate decisions, and a single pass/fail doesn't tell you which of those decisions mattered. That's the credit-assignment problem. MSCE solves it with something called reflection-weighted value backfilling: it takes the agent's own mid-task self-reflections — the little "this approach isn't working, let me try X" notes agents already generate — and uses them to distribute that one final reward back across the individual steps. Once a specific pattern of behavior consistently earns credit across many tasks, MSCE "graduates" it from a passive memory trace into a skill: a packaged object that includes not just what worked, but the evidence it's based on, when it applies, how to verify it worked, and a confidence score. Think of it like the difference between a junior engineer re-reading old ticket notes every time versus having a tested, documented internal library function they can just call.

Hands-On (~50s, ~115 words): The evidence: on EvoAgentBench, which spans five different agent task domains, MSCE takes the best or tied-best Pass@1 score in every single one. On LoCoMo, a long-horizon conversational memory benchmark, it beats the strongest existing skill-based competitor — a system called SkillFlow-Evolve — by about 2 points on judge-scored quality and roughly 1.2 points on F1. Small-sounding numbers, but they represent winning against baselines that were themselves purpose-built for skill accumulation, not generic memory retrieval. And it's training-free — none of this requires touching model weights, it's entirely a governance layer on top of what gets stored and promoted.

Takeaway (~22s, ~50 words): If you're building anything that runs multi-step agent workflows repeatedly, raw retrieved memory is a weaker foundation than verified, reusable skills with evidence attached. Worth reading if you're designing your own agent's memory layer — the credit-assignment mechanism is the reusable idea here, even outside this specific paper.

Discussion