Google AI Unveils ReasoningBank Framework That Lets AI Agents Self-Evolve Without Retraining, Boosting Performance 34%

Google AI launches ReasoningBank framework that enables AI agents to self-evolve during operation without retraining by converting past interactions into reusable reasoning strategies, delivering 34% performance improvements and 16% fewer steps across web and software engineering tasks.

• Google AI introduces ReasoningBank, a memory framework that converts LLM agent interaction traces into reusable high-level reasoning strategies, enabling agents to self-evolve during test time without retraining
• The system distills experiences into compact memory items containing actionable principles and uses embedding-based retrieval to inject relevant strategies as guidance for future tasks
• ReasoningBank combined with Memory-aware Test-time Scaling (MaTTS) delivers up to 34.2% relative effectiveness gains and 16% fewer interaction steps compared to previous memory designs across web and software engineering benchmarks