What is CODREAM?
CODREAM (Collaborative Dreaming) is a post-task reflective protocol for multi-agent AI systems in which agents that have just completed — or failed — a collaborative task jointly analyse what happened, distil the key insights, and route that knowledge asymmetrically to the teammates who need it most, permanently updating a shared experience pool without fine-tuning.
Why It Matters
Most multi-agent frameworks throw away hard-won problem-solving knowledge at the end of every task. CODREAM makes that knowledge persistent and transferable:
- Turns failures into permanent improvements: When a team of agents fails, CODREAM triggers a structured debrief. The failure is decomposed into contributing factors, each agent reflects on its own contribution, and the insights are encoded into the team's shared memory. The next time a similar situation arises, the team has already learned from it — without any human intervention.
- Asymmetric knowledge routing: Not all agents benefit equally from the same insight. CODREAM routes distilled knowledge specifically to the agent whose future performance would benefit most, rather than broadcasting uniformly. An agent that consistently struggles with type errors receives targeted type-safety heuristics; an agent performing well receives less to avoid noise.
- Zero cost: Unlike fine-tuning, CODREAM operates entirely at the prompt and memory layer. No GPU time, no retraining budget — the entire learning cycle runs on inference compute.
How It Works
CODREAM runs as a post-task phase with four steps:
- Trigger — A CODREAM session is initiated when the team completes a task (success or failure) or when a disagreement threshold is exceeded during execution.
- Collaborative reflection — Each agent generates a structured reflection on its own actions: what worked, what failed, what it would do differently. These reflections are pooled.
- Distillation — A designated "dreamer" agent (or a lightweight distillation model) synthesises the pool of reflections into a set of compact, actionable heuristics.
- Asymmetric routing — The distilled heuristics are scored for relevance to each team member's failure profile. Each agent receives only the heuristics most likely to improve its specific weak points, appended to its operational memory file.
Example
A five-agent software development team completes a sprint but ships two regressions. CODREAM triggers automatically. Agents reflect: agent-2 identifies that it skipped input validation; agent-4 notes it failed to check for null references in async paths. The dreamer synthesises: "Always validate inputs at module boundaries; add null-guard to all async handlers." These heuristics are routed specifically to agents 2 and 4's memory files. In the next sprint, both failure patterns disappear.
Relationship to Inference-Time Co-Evolution
CODREAM is one of the core learning mechanisms within the broader paradigm of inference-time co-evolution. Where co-evolution governs how agent roles and communication topology change over time, CODREAM specifically handles how experiential knowledge is captured and redistributed after each task.