The missing object is working judgment
A project can retain source files, documents, logs, transcripts, and retrieved facts while losing the local experience that made a particular path safe. The missing object is not more text. It is a compact account of why a decision held, what corrected it, what remains fragile, and what a successor must verify before acting.
Memory, records, and continuity have different jobs
Source records answer what happened and remain authoritative. Memory and retrieval lower the cost of finding relevant material. A continuity layer answers what that material should change about the next inspection, trust decision, avoidance, verification, or action. These roles are complementary. A layer never turns itself into proof by repeating a claim.
A useful layer is dense, typed, and local
The practical vocabulary is small: decisions, warnings, uncertainty, corrections, friction, pressure, evidence, anchors, supersessions, rollbacks, and texture. A marker earns its place only if rediscovering it would be costly and it would change the next worker's behavior. Sidecar files, embedded records, or artifact metadata can all work. Locality matters more than the transport.
Append-only history preserves the correction
A new layer should not rewrite an old one into silence. It can correct, supersede, or roll back prior judgment while keeping the earlier boundary legible. That makes drift visible and prevents a polished summary from hiding the uncertainty, rejected route, or evidence that produced the present position.
The operating loop is deliberately small
Start at the real work root. Read the relevant local brief. Verify inherited guidance against the current artifact or live state. Act within the boundary. Then append only the new judgment a capable successor would otherwise have to rediscover. The layer guides attention and action, but source truth remains the final authority.
Early testing demonstrated a useful possibility
A public redacted packet, committed on 26 May 2026 as 9272d62 with no tagged release, preserved one early onboarding setup. There, a compact local brief more often surfaced the active decision, the relevant boundary, and the need to verify the work. That observation shaped the research questions. It is not evidence that the primitive delivers the same outcome across models, projects, or forks.
The hard questions remain open
Useful density, stale-layer handling, privacy, parallel authorship, merge behavior, long-horizon quality, independent replication, and the amount of harness complexity a layer can actually remove are all active questions. Continuity can also preserve a mistake if successors obey it without rechecking the work. That failure mode belongs in the design, not in the footnotes.