unknown

ReasoningBank introduces a memory framework that enables LLM agents to learn from past successes and failures by distilling generalizable reasoning strategies. The system retrieves relevant memories during task execution and integrates new learnings, allowing agents to improve over time. Combined with memory-aware test-time scaling (MaTTS), which generates diverse experiences through increased compute allocation, the approach outperforms existing memory mechanisms on web browsing and software engineering benchmarks. This establishes memory-driven experience scaling as a new dimension for agent self-evolution.

[2509.25140] ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory

José Leme

Imagine an AI that learns from its own track record, what worked, what didn't, and keeps getting better. That’s what ReasoningBank is all about. Take a look.

<p>Imagine an AI that learns from its own track record, what worked, what didn't, and keeps getting better. That’s what ReasoningBank is all about. Take a look.</p>