Self-evolving agents forget earlier tasks as work drifts, new preprint finds

An arXiv preprint tests ReasoningBank, GEPA and Meta-Harness on drifting task streams and finds that more adaptation consistently costs stability.

A preprint posted to arXiv on October 19, 2026, Are Self-Evolving Agents Continual Learners?, asks whether agents that "learn" by rewriting their own prompts, memory and skills keep what they learned earlier as tasks shift. The abstract's answer is no, not reliably. If your agent edits its own system prompt or skill files based on feedback, this is the failure mode to test for.

The authors ran ReasoningBank, GEPA and Meta-Harness with frozen model weights on four task streams that drift in topic, workflow or domain, built from BIRD, BFCL and ScienceWorld. Self-evolution improved aggregate performance on most streams, but higher plasticity consistently came at the cost of stability. Methods that modify themselves make global edits to a shared prompt or workspace. Those edits interfere with unseen tasks and overwrite earlier knowledge. Similarity-based memory retrieval keeps updates more local. It still retrieves "foreign" memories on unseen tasks, and it suffers retrieval collision when similar queries need conflicting workflows.

Rivista AI's write-up reports figures from the paper. GEPA averaged a 3.7-point final gain with 4.3 points of stability loss, and ReasoningBank averaged 2.5 and 1.9. In one API-calling example, GEPA's rate of asking for missing parameters on earlier tasks fell from 29.8% to 11.2% after it learned from complete requests. @omarsar0, who flagged the paper as coming from Google DeepMind and colleagues, draws this lesson: static benchmarks hide both problems, so test on a stream that changes over time.

It's unclear whether the drifting-stream benchmark will be released as code. Two related proposals address the same problem. Self-Evolving Agent Harness proposes admitting harness revisions only after fresh, independent validation. DAIR.AI describes ReHarness, which rewrites retrieved reflections based on how a task turned out. Neither was tested in this paper.

Sources