Status: canonical scope. Accepted 2026-08-29.
Second Brain is a user-controlled, cross-agent learning and governance system. It captures evidence from agent tasks and other sources, distills durable knowledge, recalls relevant context, and turns validated user direction into reviewed guidance that can improve future agent behavior.
The system owns four planes:
- Evidence — Agent Tasks, Agent Turns, Topic Segments, decisions, insights, Correction Episodes, source ownership, and Exact Provenance.
- Learning — hybrid retrieval, task distillation, Dream Cycle synthesis, duplicate and conflict discovery, and outcome evaluation.
- Governance — inactive Steering Candidates, independent consensus, explicit approval, Authority Scope, Applicability, versioning, and supersession.
- Adaptation — bounded context packs and reviewed target adapters for
AGENTS.md, future steering files, skills, hooks, tests, lint, or CI.
Second Brain does not replace an agent harness's native memory, depend on undocumented native memory files, orchestrate arbitrary coding work, or automatically turn observed text into governing instructions. Source content is evidence; only explicit approval creates authority. Publication or enforcement is a separate, least-privilege, reviewable action.
Develop cross-cutting capabilities as Vertical Slices. Prove one Reference Integration through source evidence, durable processing, retrieval or governance, delivery, and recorded outcome before generalizing to other integrations. Codex is the current Reference Integration. A second real adapter must demonstrate what varies before a shared integration seam is extracted.
The highest-value result is not a stored memory. It is an evidence-backed improvement in what the user or a future agent can retrieve, decide, or do, with an outcome record showing whether the guidance was followed, corrected, unused, or still unknown.