Google.ai - Reasoning Bank Agent Memory Framework
TL;DR
Google AI Proposes ReasoningBank: A New Agent Memory Framework
**Core Idea:**
Google Research introduced **ReasoningBank**, a memory framework for LLM agents that enables them to **learn from both successes and failures** during task execution—without retraining. The framework converts interaction traces into reusable, high-level **reasoning strategies**, which guide future decisions and allow the agent to **self-evolve** over time.
**Key Features:**
- **Strategy Distillation:** Each experience is distilled into a compact, human-readable **memory item** (title, description, actionable principles).
- **Embedding-Based Retrieval:** Relevant strategies are injected as system guidance for new tasks, and new strategies are added after execution.
- **Failure Utilization:** Failures are turned into **negative constraints**, preventing repeated mistakes.
**Memory-Aware Test-Time Scaling (MaTTS):**
- **Parallel MaTTS:** Runs multiple rollouts in parallel, using self-contrast to refine memory.
- **Sequential MaTTS:** Iteratively refines a single trajectory, mining intermediate notes for memory.
- **Synergy:** Richer exploration improves memory, which in turn steers exploration toward better outcomes.
**Performance Gains:**
- **Up to 34.2% relative improvement** in task success compared to agents without memory.
- **16% reduction** in interaction steps, indicating more efficient task completion.
**Integration:**
ReasoningBank is a **plug-in memory layer** compatible with existing agent frameworks (e.g., ReAct, BrowserGym, SWE-Bench). It amplifies verifiers/planners by injecting distilled lessons at the prompt/system level.
**Resources:**
- [Paper (arXiv)](https://arxiv.org/abs/2509.25140)
- [GitHub (Tutorials/Code)](https://github.com/Marktechpost/AI-Tutorial-Codes-Included)
---
**In short:** ReasoningBank helps LLM agents become more effective and efficient by learning from their own experiences—both good and bad—at test time, without the need for retraining.
Comments
Post a Comment