AI Research Agents With Memory Start Faster—but Do Not Reliably Finish Better

Keeping a research notebook is not the same as knowing when its lessons apply. That distinction sits at the center of From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?, a November 2026 preprint by Jeremy Qin and colleagues. In their experiments, agents that retained state across related machine-learning tasks generally found useful solutions earlier. They did not consistently select better solutions by the deadline.

The study examines learning through accumulated working state, not changes to a model’s underlying weights. A persistent agent could retain conversation history, code, files, experimental results and execution traces. Its reset counterpart began each assignment with a fresh session and clean workspace. Within each agent, both conditions received matched tasks, data, environments and time budgets.

Testing what carries over

The authors evaluated five agents: Sonnet 4.6, Opus 5, GPT-5.4, GPT-5.6-Sol and Kimi K3. Their environment, Continual AutoResearch, covers nine task families across synthetic classification, symbolic regression, time-series forecasting and tabular machine learning. Streams contained three to five tasks, generally with 30 minutes available per task.

The tasks offered opportunities for reuse without making an earlier answer directly reusable. A forecasting procedure might remain useful while its seasonal settings needed adjustment. A symbolic-regression strategy might transfer while the equation changed. Some controlled tasks also tested whether agents would question a previously useful assumption.

Agents could inspect training and development data, run experiments and submit candidates. Private test scores were calculated after the run and were not available to guide the agents’ search. Transfer results excluded the first task in each stream, since neither condition had earlier experience at that point.

The researchers separated three outcomes: the quality of the first valid solution, the quality of the selected final solution, and research efficiency across the available time. Their efficiency measure rewards finding high-quality candidates earlier. It follows the best hidden-test score achieved so far, rather than simply counting submissions. Results were normalized and averaged with equal weight across task families.

An early advantage, not a universal final gain

All five agents showed positive average research-efficiency gains with persistent state. These ranged from 3.0 normalized reward points for GPT-5.4 to 21.3 for Kimi K3. Those values are differences in the paper’s combined quality-and-timing metric, not percentage reductions in runtime or direct changes in prediction accuracy.

Persistent agents reached their first valid solution sooner on average for every agent tested. Four also improved average first-solution quality. Yet the advantage often narrowed as reset agents continued experimenting.

Final results were mixed. Sonnet 4.6 recorded a mean final-outcome gain of 7.6 normalized reward points. Kimi K3 and GPT-5.4 were close to zero, while Opus 5 and GPT-5.6-Sol had small negative mean gains. Positive aggregate efficiency therefore should not be read as improvement on every task or every measure.

Additional analyses sharpened that distinction. When the researchers removed the delay before the first valid submission and also subtracted its starting reward, no agent retained a positive incremental-area advantage. The findings do not demonstrate a general improvement in optimization after initialization. They instead associate much of persistence’s benefit with getting a valid, sometimes stronger, starting solution into place.

Relevant memories can still mislead

A separate experiment tested written memory without continuous sessions. On the energy-systems stream, four agents began fresh sessions with either relevant insight cards from earlier tasks, cards from unrelated streams, or no cards. The cards summarized previous experiments, including evidence, confidence and conditions under which a lesson might fail.

Relevant cards changed initialization and search behavior, but did not consistently improve final outcomes. For three agents, they improved first-solution quality relative to unrelated cards while reducing the efficiency measure. Sonnet showed the opposite pattern. These results do not establish that retrieving relevant summaries reliably produces better research.

Selected execution traces illustrate possible failure modes. An agent could reuse and recalibrate a successful pipeline, but it could also become anchored to an overly specific modeling approach. Other examples showed agents replacing candidates that generalized well with alternatives that looked better on visible development data. The authors present these cases as illustrations, not a systematic causal explanation of the aggregate results.

The evidence remains limited. Most comparisons used one matched run pair; only GPT-5.6-Sol received three repetitions of the full comparison. Streams were short, and cross-agent differences could reflect their native software harnesses as well as their models. The manuscript is a preprint, so its findings remain provisional. Its central lesson is narrower than autonomous self-improvement: retained experience can improve the start of a research task without reliably improving its finish.