Google DeepMind paper finds self-improving agents forget old tasks
The fastest-learning method caused the worst forgetting, so picking a self-improvement approach by learning speed alone is misleading. Teams deploying agents that edit their own prompts or memory should isolate task-specific rules and retest prior tasks after every update. The findings apply directly to anyone building persistent agent memory on top of frameworks like GEPA or ReasoningBank.
A new Google DeepMind paper examines self-evolving agents that improve without retraining by editing their system prompt, memory, or skill files. The team tested ReasoningBank, GEPA, and Meta-Harness on task streams that changed topic, workflow, or domain, tracking gains on current tasks and losses on others. GEPA learned the most at 6.2 points per block on average but also lost the most, 4.3 points elsewhere. In one case, a later block rewrote a rule for asking users missing details, dropping a 15.5-point gain to 2.5.
New Google DeepMind paper finds that self-improving agents forget as they learn, and the method that learned new tasks fastest broke old ones the most.
Self-evolving agents improve without retraining by editing their system prompt, memory, or skill files. In real use tasks keep shifting, and each edit made for 1 task also applies to every task that follows.
The team ran ReasoningBank, GEPA, and Meta-Harness on task streams that changed topic, workflow, or domain. They tracked how much each method learned on the current task and lost on the others.
GEPA learned the most, 6.2 points per block on average, and also lost the most elsewhere, 4.3 points. After it learned to ask users for missing details, a later block rewrote that rule and its 15.5-point gain fell to 2.5.
If your agent edits its own prompt or memory, keep task-specific rules separate and retest old tasks after every update.
- arXiv: DeepSeek10-19 08:45原文
