Researchers from Google Cloud AI, Google DeepMind and Penn State have tested a system that lets AI agents carry lessons from one workflow to another. In ten controlled programming experiments, the system improved the score in nine, though the work is about refining an agent’s surrounding software, not teaching a model to rewrite its own weights.
Google DeepMind Watch analysis
What happened
The method, called Self-Improvement Memory (SIM), records why a change to an agent’s harness was made, what it was expected to do and what happened in testing. A harness is the software around a model that manages prompts, tools and task execution. The researchers’ aim is to let an agent reuse that record when the task or underlying model changes, rather than starting its optimisation from scratch.
In the reported programming experiments, SIM-backed systems scored higher in nine of ten tests, with an average improvement of 2.36 points. In a cross-model office-task test, the SIM version scored 84.75, compared with 73.42 without it. The study also found an exception: inherited experience scored 0.8 points lower than starting afresh on Gemini 4 Argon. Read 36Kr’s report.
Why it matters
An agent’s performance depends on more than the model underneath it. The surrounding software decides how it uses tools, checks its work and knows when a task is done. If useful changes and failed attempts can be carried forward, teams may spend less time rediscovering the same lessons when they switch tasks or models.
The results are promising, but they do not show that AI systems are autonomously making their underlying models more intelligent. The work described is optimisation of the harness, and the reported results vary: memory helped in the experiments, but did not win every comparison.
Our read
The interesting idea is less “AI improves itself” than “AI keeps a useful engineering notebook”. That is a more modest claim, and potentially the more practical one. The next test is whether the approach holds up across a wider range of tasks, models and independent evaluations.
What to watch
- Whether independent teams reproduce the gains across different models and tasks.
- Whether SIM reliably distinguishes reusable improvements from changes that only worked in one test.
- How performance compares with simpler ways of storing and reusing agent configurations.
Discussion spark: Should AI agents be allowed to modify and carry forward their own workflow code, or should every change be reviewed by a person?
Sources and evidence
- Google New RSI Paper: Breakthrough Agent with Built-In Memory & Autonomous Iteration for Advanced AI Reasoning – 36Kr (19 November 2026, 01:42 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind