Another RSI route emerges: model parameters stay fixed, while the Agent evolves continuously through Memory.

Beating AI News Flash: Aether AI, a causal world model company founded by UCSD Assistant Professor Biwei Huang, has open-sourced RSIAgent.

This is a recursive self-improvement framework that does not train models. The underlying model parameters remain fixed throughout, while the Agent finds tasks worth practicing on its own, actually performs them, checks the results, and then writes successful methods and failure lessons into long-term Memory. In the next round, it continues to use this experience to gradually shore up its capability gaps.

The system coordinates three Agents. The Curriculum Agent decides what to practice next, the Actor Agent actually operates the software, and the Verifier Agent independently checks the results. Exploration is divided into two steps: first broadly trying different tasks, then digging deeper into failures, hidden constraints, and edge cases. Finally, Memory is frozen and taken directly to execute formal tasks.

On OSWorld 2.0, after adding this RSI setup, the average partial score rose from 71.97% to 78.98%; Agents’ Last Exam rose from 83.75% to 84.82%. However, this is not a strict A/B comparison of the entire test suite. Only half of OSWorld tasks actually used the new results after RSI, while the remaining tasks continued to use the original scores.

It belongs to the same broad direction as Harness self-improvement approaches such as Prime Agent: model weights stay unchanged, and what keeps getting updated is everything outside the model. RSIAgent’s distinguishing feature is that it focuses improvement on Memory, then uses autonomous task generation and independent verification to let this external experience library keep accumulating.