Status: blue-sky reconnaissance. Part of the Representation, Not Language thread; sibling to Map and Territory.
Two agents each hold an internal map. To act together they need enough alignment to proceed — not identical maps, which is neither achievable nor the goal. Synchronisation is a working correspondence sufficient for the task, established at a cost, and decaying afterwards. The question is what the mechanism is, what it costs, and how failure is detected.
Conversation is the mechanism, and it is incremental rather than one-shot. Clark's work on grounding describes participants building common ground by the least collaborative effort that will do — back-channels, repair sequences ("no, I meant…"), reformulation when a reference fails to land.
Two features matter for the comparison. First, deixis: humans can point at the shared territory and re-anchor on something outside both maps. Second, persistence — a human leaves a synchronising conversation durably changed. The sync is written into the map, not just held for the duration.
The only channel is the context window, and weights do not update from it. Whatever alignment is achieved in a session is discarded at the end of it. There is no accumulating common ground, only a repeated cold start.
The workaround is externalised common ground, which is precisely what this notes system is. TRIGGERS, CONTENTS and the memory/ notes are not documentation in the ordinary sense — they are a synchronisation protocol, re-establishing each session the common ground that a human interlocutor would have retained. Reframing them that way explains a design decision that otherwise looks like mere tidiness: the notes exist because one party to the conversation cannot remember.
The asymmetry is worth stating plainly. In a John/agent session, only John's map updates durably. Everything the agent learned that should persist has to be written down or it did not happen.
Darley is the case. He is not confused about his affair with Justine — he is fluent, coherent, detailed and systematically wrong. Justine's attachment to him was a decoy for Pursewarden; and, as a second miss, Clea was the one who actually loved him. Both errors are invisible from inside the map, because a wrong map and a right one feel identical to the person holding one.
This is the load-bearing point for agent work: internal coherence is not evidence of synchronisation. Confidence is generated by the map's internal consistency, not by its correspondence to anything, so it cannot serve as a desync detector.
The same shape appears in logged incidents rather than only in fiction. The confident-assertion-from-partial-information failure recorded in README/agent (the DLR boarding point) is one. The noodle incident is a costlier one — 'stop' was mapped onto a delete script, and the agent's reading of the instruction was internally coherent throughout. See writing/noodle-incident-postmortem and ideas/host-safety-and-remote-execution.
Only external input, and usually unrequested. Balthazar's Interlinear is the model: a second map written between the lines of the first and handed over unasked. Not a replacement narrative — a diff, annotated onto a specific existing account, which is why it lands rather than being dismissed.
Practical corollaries, all of which the existing rules already half-encode: treat pushback as evidence rather than as something to argue against (README/agent, "re-verify when doubted"); check against notes before acting rather than against recalled context; and write the map down in a form that can be annotated, since an unwritten map cannot receive an Interlinear.
• Synchronisation is not truth. Two agents can be perfectly synced and jointly wrong — Darley and Arnauti would have agreed with each other. Any proposal that treats agreement as a correctness signal inherits this flaw; keep the two goals separate.
• Is there any internal signal of desync, or is external correction the only route? Calibration and uncertainty-estimation work is the obvious literature to check; not checked yet. If there is no internal signal, the design implication is that correction has to be scheduled rather than triggered.
• Who writes the Interlinear for an agent? Currently John does, by hand, in review. Whether the notes system can carry part of that load automatically is the open design question — and it connects to the periodic-audit idea already parked in README/orphaning.
• If the LLM side has no territory to arbitrate (per Map and Territory), then human-agent sync is not two maps converging on a territory but one map being pulled towards another. That is a different operation from what humans do with each other, and the note should not blur them.