The idea in one paragraph
Most agent memory stores what happened: past conversations, past trajectories, or workflows that worked. ReasoningBank stores why something worked or failed. After each task, the agent judges its own attempt and writes down the transferable lesson, such as how to approach a kind of search, or what mistake to avoid. Over time the agent builds a bank of strategies it can reuse on tasks it has never seen. The authors call this self-evolving.
How it works, step by step
- Retrieve. Before a new task, the task is embedded and the most similar memory item is fetched by cosine similarity (the default is one item). It is added to the system prompt.
- Act. The agent does the task, for example browsing a website or fixing code.
- Judge. An LLM labels the attempt as success or failure, without any ground-truth label.
- Distill. From the attempt, an LLM writes at most three memory items. Each has a title, a one-sentence description and content with the distilled reasoning steps or insight. Successful attempts give strategies; failed ones give pitfalls and guardrails.
- Store. New items are appended to the bank. The paper deliberately keeps this simple: no merging, pruning or forgetting.
| Memory type | What is stored | Uses failures |
|---|---|---|
| Trajectory memory (e.g. Synapse) | Raw past action sequences | Not as lessons |
| Workflow memory (e.g. AWM) | Routines that worked | No |
| ReasoningBank | Short strategy notes from successes and failures | Yes |
What the paper measured
The authors tested on WebArena (web tasks), Mind2Web and SWE-Bench-Verified (software issues), with Gemini 2.5 Flash, Gemini 2.5 Pro and Claude 3.7 Sonnet.
- WebArena success rate compared with no memory: 40.5 to 48.8 with Gemini 2.5 Flash, 46.7 to 53.9 with Gemini 2.5 Pro, and 41.7 to 46.3 with Claude 3.7 Sonnet.
- SWE-Bench-Verified with Gemini 2.5 Flash: resolve rate 34.2 to 38.8, with fewer steps per task (30.3 to 27.5).
- Failures help: in an ablation on the WebArena shopping tasks, adding failure-derived memories raised ReasoningBank from 46.5 to 49.7, while the workflow baseline dropped when failures were added.
- With MaTTS, spending more compute per task to produce better memories, the shopping score rose further, from 49.7 to 55.1 with parallel attempts.
The authors name their limits: they compare memory content, not memory architectures; retrieval and consolidation are kept simple; and relying on an LLM judge can introduce noise.
The misunderstanding: memory is for remembering facts
The common picture of AI memory is a list of facts about the user. ReasoningBank shows a different kind: procedural know-how. A note such as “look at the order history before searching the catalog” (our example) is not a fact about anyone; it is a lesson about how to do the work. For agents that do real tasks, that kind of memory can matter more, because it prevents the same mistake from being made twice. Where it fits among memory types: What are the types of memory in AI agents?
What this means for a team
A team generates exactly the material ReasoningBank learns from: corrections, rejected drafts, bugs that came back. Most of it disappears into chat history. The lesson from the paper is to capture the rule behind the correction, in a short form an agent can retrieve, and to keep the failures. Two cautions carry over. An automatic judge can be wrong, so in a team a human confirmation on important lessons is worth its cost. And a bank that only grows needs cleanup; see How should an AI agent manage its memory? Practical steps are in How to make AI agents learn from corrections.
Where Bombos fits
ReasoningBank lets one agent learn from its own attempts. Bombos is a plugin for the team version: a shared memory layer underneath Claude, ChatGPT, Copilot, Grok, Codex, Cursor and your own agents that learns how your team works, so what one tool learned is there for the others. With every answer you see what it rests on. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.
Related questions
Who wrote ReasoningBank?
Siru Ouyang, Jun Yan and co-authors, most of them at Google Cloud AI Research, with contributors from Yale and the University of Illinois Urbana-Champaign. It is arXiv 2509.25140 and was accepted at ICLR 2026.
How does the agent know whether it succeeded?
An LLM acts as judge and labels each attempt as success or failure, without ground-truth labels. The authors note this can add noise when a task is ambiguous or the judge errs.
What is MaTTS?
Memory-aware test-time scaling: the agent spends extra compute on a task, either by running several attempts in parallel and comparing them, or by refining one attempt step by step, and uses those extra experiences to write better memory.
Is the code available?
The authors link a repository at github.com/google-research/reasoning-bank.
