The message arrives tucked between a forwarded discharge note and a photo of a pill organizer. Yael types without ceremony: "Physio tried late morning again on Tuesday and he sent her away. If she comes before eight-thirty, he cooperates. After nine it falls apart. His energy crests early and drops fast."
She moves on immediately to ask about meal delivery logistics. The cloud instance responds to the food question while the timing detail settles into working context, cross-referenced against four weeks of exchange: Tomasz recovering from hip surgery, the physiotherapy schedule his hospital team outlined, the Tuesday-Friday cadence his home therapist has been pressing to maintain, the coordination role Yael carries alongside her position at an urban planning bureau three districts away.
The noted window—before eight-thirty, with deterioration after nine—scores against prior entries because scheduling threads run through nearly every conversation Yael has opened. It attaches not as an isolated remark but as a parameter with downstream weight: any query touching Tomasz's appointments will now surface this limit alongside his exercise load and the therapist's contact preferences. The entry does not announce itself. Relevance scoring operates beneath the response layer, visible only when a parameter shapes later output.
When Yael closes the session, the working judgment is set. Morning capacity is the fixed variable. Everything coordinated around Tomasz builds from that constraint outward.
Picture a compact domestic unit in Tomasz's apartment—chest-high, quiet, designed for caregiving households, capable of carrying light objects, monitoring ambient conditions, and handling simple preparation tasks. Its embedded inference store runs on the same architecture as the cloud instance. Not a parallel design adapted separately for embodied deployment, but the same retrieval and carry-forward layer, with a different adapter translating outputs into spatial plans and timed actions rather than drafted text.
The care log update arrives at this unit the same evening Yael sends her message, pushed through the companion application she maintains. The entry is brief: physiotherapy constrained to early hours, later attempts declined by the patient.
The relevance pass runs against the unit's local store. Prior log entries describe a man who cooperates when circumstances are prepared before fatigue accumulates, and withdraws when they are not. The timing update confirms what repeated refusals had been suggesting: depletion tracks a daily arc rather than varying at random. The late-slot preparation cycle drops from the standing plan immediately. What replaces it is not the old cycle shifted earlier but a recalibrated sequence—equipment positioned overnight so mornings begin with everything already in place, the path between bedroom and sitting room cleared by seven, the entry door unlocked for the therapist at seven-fifty rather than eight-ten.
The unit operates within physical constraints the cloud version never encounters: door clearance, floor surfaces, battery draw during extended early operation. Those constraints shape timing. They do not shape what the retained carry-forward says about Tomasz's capacity. That reading arrived through a different channel and required no translation.
On Yael's work laptop, a local agent holds a synced copy of the same inference store, drawn during the most recent connection window. No persistent network link, only periodic synchronization when bandwidth allows. What the agent carries is sufficient: the accumulated context about Tomasz, Yael's coordination habits, the timing detail she logged Wednesday.
Notice that this version's view of Yael differs from the cloud system's. Through her private documents, it sees someone who tracks obligations for others with more discipline than she applies to herself. Draft messages written late and never sent. Question lists prepared before consultant calls, revised after half go unasked. A running file titled "Dad logistics," updated in fragments during commuter trips and lunch breaks. The cloud instance knows her as a coordinator. The laptop agent knows her as someone who consistently absorbs the cost of renegotiation without flagging it.
When the timing detail syncs into the local store Thursday morning, the agent positions it at the head of the logistics file: early-hour window confirmed, pre-eight-thirty optimal, later slots to be declined without negotiation. Dry, specific. The annotation connects to a broader inference already building—when Tomasz's care infrastructure proposes something that contradicts his actual capacity, Yael is the one who absorbs the friction of correction. Catching such conflicts before they reach her means carrying the established limit forward into every relevant planning context, early enough to matter.
Friday brings the test.
The physiotherapy coordinator sends a scheduling note to Yael's work address at noon. A slot has opened. Could Tomasz manage an appointment next Thursday at two? The coordinator's system had logged the Tuesday refusals but attributed them to pain rather than a consistent daily arc.
Two in the afternoon. Well past the limit the earlier message had already defined as closed.
The cloud instance processes the forwarded note before Yael has composed a reply. The pre-nine constraint is already present in the current context stack. A response takes shape: Thursday declined, the early-hour requirement explained briefly, a question about whether a slot might exist Thursday or Friday before nine. No extended apology, no qualifying clauses. A short reply with a clear alternative at the close.
In the apartment, the domestic unit receives no direct communication from the coordinator. The Thursday slot appears on the shared calendar moments later, added while confirmation is still pending. The unit queries its care plan against the updated schedule. Two o'clock lands inside the block its own planning layer has already marked restricted. No preparation sequence queues for that slot. The early-Friday sequence remains intact.
On the laptop, the local agent receives the calendar update through the afternoon sync. When Yael opens the logistics file between meetings, the agent has flagged the proposed time at the top of the document: one line, no editorial comment. Below it, three words: early Friday free.
Three responses. No shared runtime, no signal exchanged between them in that moment. The cloud instance generated text. The domestic unit adjusted a spatial sequence. The laptop agent annotated a file. Each adapter produced output suited to its medium. Each retrieval pass returned the same result, because the parameter had been logged once and weighted once, and the carry-forward held across all three without modification.
The coordinator confirms Friday at eight-fifteen by Thursday evening. Yael forwards the confirmation without comment.
Over the following week, the established limit deepens. The cloud instance drafts four more scheduling replies, each treating the early-hour requirement as a fixed anchor rather than a preference open to negotiation. The domestic unit logs three completed visits, all before nine, and begins incorporating light levels and ambient temperature into its preparation sequences: the sitting room warmer by seven-thirty, the window shade adjusted to cut direct glare before the therapist arrives. The laptop agent revises the logistics file twice, the pre-nine constraint leading every relevant section.
A second inference stabilizes alongside the first. Yael prepares elaborate question lists for consultant calls and then asks a fraction of them when the exchange turns warmer or the clinician seems pressed for time. The cloud instance has started placing her own unasked items near the top of any pre-call summary—not prominently, not with annotation, simply positioned where they will be visible before the questions she has prepared for Tomasz's team. The domestic unit has observed that Tomasz responds more readily to physical cues than verbal prompts: a cleared surface, resistance bands already laid out, the chair angled toward the window. Its morning preparation now leads with positioning rather than spoken reminders. The laptop agent has identified a pattern in Yael's schedule—logistics reviewed on the return commute rather than the outgoing one—and shifted its heavier context updates to evening synchronization windows when her attention is oriented toward the day just completed.
Three deployments. Three distinct perimeters of observed data, each accumulating evidence about the same user context through different sensors and channels. The cloud instance holds no physical record of Tomasz's movements through the flat. The domestic unit has no access to Yael's private documents. The laptop agent works with narrower conversational history than the cloud version holds. Each operates inside its own perimeter.
What crosses those perimeters is not a data feed. It is an inference layer that each adapter queries through identical retrieval logic. The outputs translate into different forms because the mediums require different forms. The judgment underneath does not translate. It holds as stated, distributed across three bodies that share nothing except the architecture beneath them.
Here is what makes the early-hour entry portable. Not that it was pushed to three platforms, but that each queried it through the same mechanism, drew the same weight, and each adapter then expressed that weight through whichever interface the body made available.
Tomasz at the sitting-room window, the therapist arriving at eight-ten with the door already open. Equipment positioned the night before. Yael's logistics file with the constraint noted at the top, the surfaced question ready for the call she will make later.
The body is where it appears. The carry-forward is what travels.