Wearable memory assistants capture conversational audio all day and distill it into a memory layer — summaries, entities, commitments. Today the category is evaluated on transcription accuracy and recall quality: can it answer questions about the past? That leaves the harder claim untested — whether the memory layer models the wearer, or merely indexes them.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
The ambient-AI stack is converging along one ladder: recording → structured memory → longitudinal evidence → target-specific predictive capability → validated assistance, with the predictive-capability rung measured by TargetSpace. Memory devices are the canonical memory systems rung: continuous capture already gives them the raw material of longitudinal context, but the derived layer is built and marketed for recall, not for prediction.
Current rung: structured memory — store, summarize, and retrieve what the wearer said and heard. Next rung: longitudinal context — treating the accumulated record as evidence about the wearer's future transitions, a claim that must be tested rather than assumed.
In the flagship TS-Personal track, memory devices contribute L2 evidence — conversational audio — over an L0–L1 digital-exhaust floor. The derived memory layer itself (summaries, entities, commitments) is treated as a model component under test, not as ground truth: whether it adds forecast skill beyond the raw tiers is exactly what the evidence-tier ablation measures. Memory is not understanding; retrieval is not understanding — the protocol asks whether remembered context improves calibrated forecasts of what the wearer does next.
Run the same sealed forecast tasks twice — once with the memory layer enabled, once with it disabled. Each task seals, before the outcome exists, a probability distribution over a discrete answer space: for example, whether a commitment the wearer stated aloud this week is completed, deferred, cancelled, or replaced by its deadline. Outcomes resolve deterministically under pre-registered rules; forecasts are scored with proper scoring rules against the R1 population-prior baseline and the R2 own-routine baseline, with calibration reported alongside skill.
The memory layer earns target-specific credit only if the enabled condition adds calibrated skill over R2 and that lift collapses under the permutation specificity gate — the same forecasts matched to the wrong wearer. If skill survives permutation, the layer captured generic patterns, not this wearer.
Scope. This measures calibrated, target-specific forecasting skill about future observable states — a bounded, product-facing notion of personal intelligence. It makes no claim about inner life, true intent, or what the wearer “really” means.
A feature on/off A-B over the same sealed task set: memory layer enabled vs. disabled, identical targets, identical horizons. One switch, one frozen battery — the difference in calibrated skill is the memory layer's measured contribution.
Raw audio and the memory store never leave the company. Only sealed forecasts, resolved outcomes, and aggregate metrics are exported — the protocol audits skill without ever seeing the wearer's data.
A leaderboard row: Skill vs. R1, Skill vs. R2, calibration, permutation result — reported per evidence tier, so the marginal value of conversational audio and of the derived memory layer is visible separately.
Does the memory layer improve prospective forecasts, or only question answering about the past? The two are routinely conflated; the on/off A-B under the permutation specificity gate separates them.
Which stored memories carry predictive information about the wearer's future transitions, and which are inert? Per-memory ablation over sealed tasks could rank the store by forecast value rather than by recency or salience.
What memory-decay or consolidation policy maximizes forecast skill per stored byte? Skill-per-byte gives compression and forgetting an objective the category currently lacks.
Does remembering what the wearer said they would do add lift over what they were observed doing, given the intention–behaviour gap? Self-report is biased but not dismissable — it is an auxiliary channel whose marginal value should be measured.
How should a memory layer represent uncertainty about what it stored — misheard audio, wrong speaker attribution, stale summaries? Calibrated forecasting penalizes overconfident memories; the store may need confidence of its own.