Google's ReasoningBank: distilling strategies from trajectories turns LLMs into self-improving agents

solyarisoftware · x · 2026-08-26

A new Google paper proposes ReasoningBank memory extraction and memory-aware test-time scaling (MaTTS), establishing the first closed-loop architecture for memory-driven agent self-evolution.

Pipeline: streaming task query → semantic memory retrieval → memory-guided rollout → LLM-as-a-judge signal → distill strategies from failure/success trajectories → consolidated memory bank.

Key ideas:

Strategy-level reasoning distillation replaces lossy raw trajectory dumps with structured reasoning items (title, description, decision rationale) that transfer across unseen websites and codebases;
Counterfactual learning from failures;
Demonstrates failure-driven emergent strategy evolution across web and coding benchmarks.

Original post →

Google's ReasoningBank: distilling strategies from trajectories turns LLMs into self-improving agents

More from coding & agent