A wearable already has a memory strategy: keep almost nothing locally, send the rest to the phone or the cloud, clear the session between glances. That strategy spends the budget it was trying to save, because the radio is the most expensive thing the device does. A bounded field keeps the day's context resident inside the envelope the hardware was designed around: a footprint you set at build time and that does not move for the life of the SKU.
A wearable manages memory by not keeping it. The on-device KV cache is sized for one short exchange against whatever SRAM or LPDDR the SoC was specified with (a Snapdragon AR1 or W5+, an Ambiq Apollo, an nRF54), and anything longer goes over BLE to the phone or over Wi-Fi to a server. The session clears between glances. Longer-term context, where the product has any, lives server-side and comes back as a retrieved snippet on the next request.
That strategy spends the budget it was built to protect. Every offload is a radio wake, the most expensive operation in the power envelope per useful bit, and it is paid twice: once in milliamp-hours and once in heat against a skin-contact surface with a fixed ceiling. It puts the wearer's day on someone else's server to make recall possible at all. And it fails outright on a plane, in a basement, or anywhere the link drops, which is where a wearable is supposed to be better than the phone.
The deeper problem is that a retrieval-backed memory system has a footprint that moves with usage, and a hardware team cannot specify against a number that moves. Memory is the one line in the power model that the industrial design cannot absorb late. It freezes at tape-out, alongside the battery and the enclosure, and everything after that is a negotiation against a fixed ceiling.
SCOPE: schematic. Measured footprint at your silicon and power target comes out of the pilot, at the fidelity setting you choose.
Device-class behavior below is projected from measured datacenter results and confirmed against your silicon in a pilot.
A full day of glances runs to hundreds of thousands of tokens. The field carries that day inside a footprint you fix at build time, on the silicon you already specified.
SCOPE: measured on a single consumer-grade NVIDIA Blackwell GPU, on a single unbroken conversation of 1,900,000 tokens across 9,359 turns, 14.5× the base model's trained context, with the resident memory flat from 50,000 tokens through to the final turn and total footprint ~1,370× at 1.9M tokens under a standard cache. The ratio grows with length and is only meaningful at a stated depth. Reproduced on 3 independent conversations. At the load a full day of glances puts on context, that is several days of continuous use in one unbroken session. Your device-class footprint, and the transmit energy it displaces, are projected from these results and measured in the pilot.
Continuity stops costing a radio wake. The day's context is resident, so the transmit energy it used to consume returns to the battery and to the thermal budget.
Memory footprint becomes a number in the industrial design, fixed at the same stage as the battery and the enclosure, rather than a variable that moves with how much the wearer talks.
Recall runs against local state, so the wearer's day does not have to sit on a server for the assistant to remember it.
The assistant stops failing partway through. Before we bounded the read, our own runs died at roughly 285,000 tokens: output collapsed to about a dozen tokens a turn and stopped. At the load a full day of glances puts on context, that ceiling falls inside the first day or two of continuous use. Nothing like it appears at any depth we have run since.
SCOPE: on-device behavior depends on your silicon and power target, and is projected from results measured on a single consumer-grade NVIDIA Blackwell GPU. A paid pilot returns the measured footprint and the transmit energy it displaces, at your chosen fidelity setting.
A full day of glances carried as one resident structure inside the footprint the industrial design already committed to. Context from the morning is read at the same resolution as context from the last interaction. In our deepest published run a thread from the opening was picked back up ~1.6 million tokens later, never re-introduced by name — the morning read as sharply at the end of the session as the last thing said.
The current interaction at full fidelity, over an instruction block the field cannot overwrite. Sized for response latency rather than conversational depth, because a glance is usually one turn.
The footprint is a number you set at build time, against the SRAM and DRAM you already specified. It is the same number on day one and at the end of the SKU's life.
Footprint held flat across a full day of use, at your chosen fidelity setting, against your actual SRAM and DRAM budget.
Transmit energy displaced, counted as offloads avoided and converted to milliamp-hours and to hours of runtime on your battery.
Recall with the radio off, against the same workload served from the cloud.
Glance-to-response latency with no round trip, against your current offload path.
SCOPE: this page states direction, not magnitude. Measured numbers for your workload come out of the pilot, with the configuration that produced them.
Bounded memory on its own is not hard. A rolling window gets the same footprint number tomorrow by dropping everything older than the window, and on a device where the whole point is that the assistant knows the wearer's day, dropping it is the failure. What is hard is holding the footprint flat while the whole day stays available and the model is still reading the start of it at the end. That is the result we publish, reproduced across independent conversations, and it is what a pilot reproduces against your silicon.
The field is held to a budget you fix before you build. It runs on the model you already ship, with no retraining and no recalibration. Patent pending; the implementation stays private.
A scoped four-week pilot: you bring one workload and a memory budget, we return a measured comparison against your current cache: footprint, recall, and the configuration that produced both.