CoDream (collaborative dreaming) lets a team of AI agents learn from a failed task without retraining the AI model underneath. After the failure, the agents talk through what went wrong, write the lesson down as a short rule and hand it only to the teammates who are weak at that kind of task, like a debrief after a lost deal. It comes from a research paper of May 2026.
Where does CoDream come from?
It is the core of EVOCHAMBER, a framework that researchers from Oregon State University, Johnson & Johnson and other institutions put on arXiv on 11 May 2026 (Zhang et al.). EVOCHAMBER keeps a pool of AI agents improving while they work through a stream of tasks, without changing the model's weights. The authors call this test-time co-evolution and run it at three levels: each agent keeps its own lessons, a team of three is assembled per task, and at the population level CoDream moves knowledge between agents while other operators add, copy, merge and remove agents.
How does a CoDream session work?
A session starts when a team fails: the average score falls below a threshold, or the members disagree. The three team members then go through five phases:
- Reflect: each member diagnoses its own attempt.
- Contrast: failing members are paired with successful ones to find what the successful approach did differently.
- Imagine: those differences become candidate strategies, tagged with the types of task they might apply to.
- Debate: the members criticise each other's proposals and drop the weak ones.
- Crystallize: the surviving proposals become structured insights, each scoped to one task, a subdomain or all domains.
Each insight is written into every agent whose competence on that type of task is below the median of the pool. Strong agents produce knowledge and weak agents receive it, so specialists stay specialised. Sending every lesson to every agent, as the earlier method MemCollab does, dilutes that specialisation, the authors argue. The lessons end up in the agents' experience archives, and the most similar ones are added to an agent's prompt for later tasks.
What did the tests show?
With the open model Qwen3-8B, removing CoDream cost more than removing any other part: on a stream of 600 tasks from six domains, accuracy fell from 87.1% to 76.3%, with the largest drops on HotpotQA (89.5% to 57.2%) and DROP (86.0% to 48.0%). In a smaller check on 30 math tasks, the pool of agents without CoDream scored exactly as well as a single agent (63.3%); with CoDream it reached 70.0%. The full system reached 63.9% on hard math problems against 48.4% for the best comparable method, MemCollab: 32% better in relative terms, a margin that applies to math only.
What does it cost, and where are the limits?
- More compute. The whole system uses roughly 3.6 times the inference cost of a single agent, which the authors call possibly prohibitive where speed matters. It was still more accurate than five-way majority voting at 72% of that method's tokens.
- The evidence is narrow: two model families, benchmark streams of up to 600 tasks and fixed thresholds. Longer streams and lessons that go out of date are left to future work.
- The paper notes that in its implementation most insights were classified as cross-domain; routing to the weak agents on one type of task applies when an insight is judged less transferable.
Example
One insight that CoDream produced after a CodeContests task, quoted in the paper: "When input validation involves numerical ranges and potential overflow, use bounded arithmetic with explicit saturation: clamp intermediate values to the valid range [min_val, max_val] using min(max(value, min_val), max_val) before further computation." The code is on GitHub under the MIT licence. The post How can multi-agent AI systems learn from their failures without retraining? covers EVOCHAMBER as a whole.