How Much Memory Does Your Agent Actually Need?

Agents can learn from their own past work by extracting behavioral guidelines and reusing them at inference time, a process that requires no model weight updates. Research across eight models found that the optimal amount of memory to give an agent depends on the model's capability level, not simply on accumulating more guidelines. Strong models with available capacity benefit from receiving a complete set of guidelines, weaker models perform better with a selective subset retrieved per task, and already-saturated models show no measurable improvement. The practical implication is that memory should be calibrated as a dose tailored to each model rather than treated as a feature to uniformly enable.