MESA: the agent selects memory for the task instead of reading everything
Sh0ny
12 августа 20262 min read
In short
On long tasks it hurts an agent to load the whole history automatically: the extra context adds noise and burns tokens. MESA proposes assembling memory around the specific question, and in testing it showed both higher accuracy and a smaller body of evidence.
On long tasks an agent is hampered not only by forgetfulness but by a surplus of memory. Read the whole history of actions, observations and reasoning and a great deal of noise enters the context; pick only one source and you may lose an important part of the answer.
That trade-off is exactly what the MESA work examines. Its main conclusion sounds practical: the best memory configuration usually lies in the middle — not a single structure and not a full merge of all of them, but a small set chosen for the specific query.
For each trajectory the authors build five complementary memory representations. MESA then learns from feedback on the final answer: it picks the representations it needs and merges the information drawn from them, leaving the answer model itself frozen.
That is an important difference from the simple rule of "load more context". The system has to decide first which kinds of memory are needed right now. To learn that selection it uses harness optimisation with a search that accounts for previous results, plus UCB planning that balances exploring new options against exploiting ones that have already worked.
On AMA-Bench, MESA beat the strongest baseline by 8.5% while using 41% fewer evidence tokens than the variant that reads every structure. So saving context here does not look like a price paid in quality — by the reported results it comes together with it.
But there is an important caveat: the results come from a single benchmark, and the abstract does not disclose which five memory representations are used or how expensive the selection mechanism itself is to train. So this is not yet universal proof that every agent needs such a routing layer. It is rather a strong argument against one identical memory strategy for all queries.
If your agent already stores long histories, would you first optimise the volume of context or teach the system to choose which memories are worth reading at all? Source: cs.AI updates on arXiv.org