Memory for LLM evolution only helps if it may stay silent

The main result of this work is not that LLMs have been taught to remember. What matters is something else: past experience really does speed up the search for new algorithms, but only if the system knows when not to use it.

ε-MemEvo stores not source code but short descriptions of successful strategies — tactic memories. That format transfers between tasks with different APIs and checking functions: the system does not need to copy an old solution literally.

But memory easily turns into a hindrance. So before the discovered experience is added, an adaptive injection gate operates: it decides whether to let the hint through and how heavily to lean on it. While things are improving the search more often chooses skip, and on an early or late plateau it may switch to hint. In other words, memory here is not a base of mandatory instructions but conditional advice.

Across eight optimisation tasks ε-MemEvo showed a better AUCC than AdaEvolve on every one of them when using GPT-5: the average relative gain was 8.7%, and early convergence improved by an average of 9.4%. The extra computational cost is under 1%.

The limitations are substantial too. This is a result on eight benchmarks, and the headline figure comes from one backbone — GPT-5; the Leave-One-Out protocol excluded the target task's memory but does not turn the experiment into proof of the method's universality. Besides, ablations showed that naive memory injection can fail catastrophically, whereas the adaptive gate stayed safe across the five tasks tested.

The practical lesson for agent systems is fairly down to earth: collecting a history of successful actions is not enough. You need a separate mechanism that judges a memory's relevance during the search itself. Otherwise "memory" will be not experience but one more source of confident errors.

If you already use memory in an AI agent, does it more often help it find something new or stop it escaping an unsuccessful pattern? Source: cs.AI updates on arXiv.org