journal

Cross-round memory for meta-agent workflows is maintained by storing per-round hypotheses, outcomes, and gating decisions in a structured journal. This records what was tried, what landed, and what got reverted so future rounds can avoid repeating doomed hypotheses.

Experiment hypotheses and outcomes require per-round YAML frontmatter containing round, timestamp, hypothesis_id, levers, gating_outcome, gating_attribution, and predicted_affected fields, followed by a free-form prose body detailing the results.

Meta-agent context is summarized for quick recall by exposing an indexed CONTEXT.md file. The orchestrator generates this file to provide quick context on previous rounds without parsing the entire journal history.

Yes, the journaling approach supports audit and learning by identifying failing hypotheses across rounds. Storing gating outcomes and attributions explicitly guides future rounds to avoid repeating doomed hypotheses.

To update the workflow journal after each evolution, append one new round section containing the required YAML frontmatter and prose body. This continuous appending builds the cross-round memory needed to guide future agents.

Using YAML frontmatter for workflow orchestration memory requires strict adherence to specific fields like gating_outcome and hypothesis_id. Missing or malformed frontmatter disrupts machine-parseability and prevents the orchestrator from exposing valid CONTEXT.md summaries.