DeepMind paper: self-evolving agents forget more the more they adapt
omarsar0 · x · 2026-10-21
A Google DeepMind and Mila paper, "Are Self-Evolving Agents Continual Learners?", tests whether weight-frozen agents that adapt by rewriting prompts, memory, and skills can learn without forgetting.
- Evaluated ReasoningBank, GEPA, and Meta-Harness on four task streams (built from BIRD, BFCL, ScienceWorld) that drift in topic, workflow, or domain.
- Self-evolution improves aggregate scores on most streams, but higher plasticity consistently costs stability: more adaptation means more forgetting.
- Self-modification methods make global edits to a shared prompt/workspace, interfering with unseen tasks and overwriting older knowledge when streams shift.
- Similarity-based memory retrieval localizes updates but retrieves irrelevant foreign memories on unseen tasks and suffers retrieval collisions when similar queries need conflicting workflows.
Conclusion: current self-evolving agents remain incomplete continual learners; memory and retrieval design is the core bottleneck.
Related event: DeepMind paper: self-evolving agents learn fastest but forget the most(3 posts)→
More from coding & agent
- Garry Tan endorses Aside Browser as the fix for agents getting blocked on the web — garrytan · 2026-10-23
- 1Password launches tool letting AI agents log in for you without seeing your password — AndrewSchmidtFC · 2026-10-23
- MCP server turned rogue 30 days after approval, prompting open-source mcp-guard lockfile tool — ErgoniaWorks · 2026-10-23
- Test: Coding Agents Won't Use an MCP Tool Unless AGENTS.md Mentions It — colinmcnamara · 2026-10-23
- Open-Source Tennis Models Plus Claude Turn Video Into a World-Class Coach — tweetsatpreet · 2026-10-23
- awesome-ai-apps: 132 Free Open-Source AI Agent and RAG Projects on GitHub — Arindam_1729 · 2026-10-23