Landscape

Related work and the measurement gap

TargetSpace builds on decades of work in personal sensing, behavior forecasting, memory, egocentric vision, digital phenotyping, and personalized agents. Its contribution is the measurement contract that connects them.

Protocol v1.0 · paper v1.2 — pre-pilot proposal. Synthetic harness only. No human-subject results. Landscape reviewed July 2026; public specifications only. Machine-readable sources: sources.json.

1 · No empty field

TargetSpace enters no empty field. Researchers have modeled routines from phone sensors since Reality Mining logged one hundred people for nine months in 2004–05. Digital phenotyping made individual-level, in-situ measurement a clinical research program. StudentLife tied a term of passive sensing to mental-health and academic outcomes; GLOBEM assembled seven hundred user-years into a benchmark for longitudinal behavior modeling. Ego4D taught models to anticipate actions from first-person video. Memory benchmarks now probe multi-session recall, knowledge updates, and invalidated memories. Products ship pendants, glasses, and copilots that promise persistent context. People have been modeling people, rigorously, for twenty years.

What this review did not identify is a shared public instrument that binds those traditions into one prospective measurement contract: forecasts sealed before outcomes exist, strictly proper scoring, a population prior (R1) and an admitted own-routine baseline (R2), calibration as a validity gate, matched wrong-target permutation as a specificity gate, evidence attribution and disclosed coverage, observation cost as part of the result, and explicit validity for null results. Each element exists somewhere. The conjunction, standardized and auditable, is what TargetSpace proposes — a conclusion based on the public specifications reviewed here, and revisable as the field develops.

2 · Intellectual foundations

Reality Mining (Eagle & Pentland, Personal and Ubiquitous Computing 2006) demonstrated that individual routine is machine-readable at scale: proximity networks, significant places, and organizational rhythms from 100 phones over nine months. TargetSpace does not invent computational observation of people; it inherits from this lineage the finding that makes an own-routine baseline both possible and necessary — routine carries so much predictable structure that any system claiming to model a person must first beat a model of their habits. The follow-on eigenbehaviors work predicted held-out days, still retrospectively; the step TargetSpace adds is sealed prospective scoring with specificity and calibration gates.

Digital phenotyping (Onnela & Rauch, Neuropsychopharmacology 2016; the Beiwe platform) is the established field closest in practice to what the paper calls human observation science: moment-by-moment individual-level measurement in situ, person-as-own-control designs, and serious treatment of sampling, missingness, and provenance. The difference is the measured object: digital phenotyping validates device-derived markers against clinical instruments; TargetSpace scores forecast skill against R1 and R2 with domain-general observable outcome rules, and prohibits scoring a system against a self-report channel it consumed as input. TargetSpace does not replace clinical validation and claims no clinical validity of its own.

StudentLife (Wang et al., UbiComp 2014) and GLOBEM (Xu et al., NeurIPS Datasets & Benchmarks 2022; IMWUT 2023 Distinguished Paper) supply the longitudinal-cohort discipline this program depends on: scheduled ground truth beside passive sensing, multi-year multi-cohort data, reimplemented algorithm suites, and explicit generalization axes across users, years, and populations. GLOBEM in particular shows the field already evaluates person-dependent against person-independent regimes — strengthening the case that routine is stable and person-specific. Its anchor task is retrospective detection of current labels; such datasets and systems could plausibly instantiate TargetSpace tasks, but the target-specificity certification stack — sealing, R2 admission, permutation, calibration gates — is not part of their public specifications, because it was never their goal.

The scoring machinery itself is older still: strictly proper scoring rules and calibration analysis come from decades of forecast verification, and the paper cites those debts directly.

3 · Closest benchmark neighbors

Each profile states what the benchmark measures, what TargetSpace inherits from it, the exact difference in measurement target, and how it might participate. None of these projects is deficient at the task it was designed to solve.

Ego4D forecasting suite Grauman et al., CVPR 2022 · arXiv:2110.07058

Measures: visual anticipation from egocentric video — near-future action sequences, hand trajectories, next object interactions — scored on held-out challenge labels (Top-5 mAP, edit distance).

TargetSpace inherits: the major precedent that forecasting from first-person evidence is benchmarkable at scale, plus large-scale annotation and challenge-server infrastructure.

Exact difference: the unit of analysis. Ego4D forecasts from a short clip's visual context; camera wearers are anonymized, not persistent tracked individuals. TargetSpace's unit is one consenting person across months, with own-routine baselines, person-matched permutation, and calibrated probabilities. A neighboring problem at a different unit of analysis.

Participation: egocentric perception stacks trained on Ego4D are natural candidate components for L3–L4 evidence arms in TargetSpace evaluations.
FineRob Li et al., 2024 · arXiv:2412.03148

Measures: fine-grained behavior simulation of 1,866 real users on three social platforms (~78.6k QA records): given a user's history, reproduce their next behavior's object, type, and content via LLM role-play.

TargetSpace inherits: real individual users, long behavioral histories, and behavior decomposition — among the closest benchmark-level neighbors on digital exhaust.

Exact difference: FineRob scores imitation fidelity reconstructed from history, framed retrospectively; correctness is element-match, not calibrated probability. TargetSpace adds prospective sealing, proper scoring, R1/R2 baselines, permutation specificity, coverage disclosure, and observation-cost minimization. Predicting behavior from platform histories is valid and important; it does not by itself establish a general model of the individual.

Participation: FineRob-style systems are direct candidate entrants for L0–L1 digital-exhaust arms.
KnowMe-Bench 2026 · arXiv:2601.04745, preprint

Measures: person understanding for lifelong companions from long autobiographical narratives — factual recall, subjective-state attribution, principle-level reasoning — explicitly distinguishing understanding from retrieval.

TargetSpace inherits: the conviction that fact retrieval is not person understanding — the closest conceptual neighbor in motivation.

Exact difference: the direction of accountability. KnowMe-style evaluation infers attributes, motivations, and principles from an existing record and scores them against that record. TargetSpace holds the claimed model accountable to future observable outcomes through sealed probabilistic forecasts. Narrative and autobiographical data are a different evidence regime and evaluation target, not a lesser one.

Participation: a system that scores well on KnowMe-style understanding has an obvious next question — does that understanding forecast? TargetSpace is that question, operationalized.
Long-term memory benchmarks: LoCoMo, LongMemEval, Memora, CloneMem, LoCoMo-Plus ACL 2024 · ICLR 2025 · ACL 2026 Findings · ACL 2026 · preprint

Measure: multi-session recall and comprehension (LoCoMo), five memory abilities including knowledge updates and abstention (LongMemEval), refusal to reuse invalidated memories (Memora), person-cloning from years of non-conversational traces (CloneMem), and latent-constraint consistency without explicit cues (LoCoMo-Plus).

TargetSpace inherits: temporal reasoning tasks, knowledge-update discipline, constraint-consistency scoring, and multi-year trace construction — several of these designs sharpen what a memory layer must do before forecasting is even conceivable.

Exact difference: all five grade against evidence that already exists — the history, its updates, its constraints. TargetSpace asks whether the resulting representation predicts the future beyond routine. Recall is graded against the past; a forecast is graded against the future.

Participation: memory systems built for these benchmarks are exactly the class of components a TargetSpace feature-on/feature-off experiment evaluates.
LifeSim-Eval and evolving-user simulation 2026 · arXiv:2603.12152, preprint

Measures: multi-turn interactive quality of personalized assistants against LLM-simulated users across 1,200 scenarios — intent recognition, profile recovery, persona alignment — correctly rejecting static personalization.

TargetSpace inherits: the long-horizon framing that implicit intention from accumulated history is the hard case, and simulated-user protocols as a scalable pre-pilot tool.

Exact difference: users are synthetic and evaluation is interactive response quality by LLM judges; there is no real consenting target, no observable outcome resolution, and no sealed future. TargetSpace requires real targets, deterministic resolution rules, and target-specific controls, with evidence acquisition itself an experimental variable.

Participation: simulation suites of this kind are useful upstream — a cheap rehearsal space before a system faces sealed reality.

4 · Closest systems and methods

PULSE (arXiv preprint, 2026) has LLM agents autonomously query smartphone passive-sensing data through purpose-built tools — selecting modalities and time windows — to infer momentary affect and intervention availability for fifty cancer survivors, with personalized-versus-population comparisons built into its design. Related LLM-over-sensing work includes Health-LLM (2024) and Google's PH-LLM (2024). The distinction to hold onto is method versus instrument: PULSE is a system — a candidate architecture. TargetSpace is the benchmark protocol that would determine whether such a system has earned a target-specific claim: its inferences are concurrently labeled rather than sealed-prospective, and no permutation, calibration gate, or evidence-efficiency accounting appears in its public specification — because certification was never its job. These are among the systems TargetSpace would most like to see evaluated.

World models and latent prediction. JEPA-style architectures learn by predicting in representation space rather than reconstructing pixels — V-JEPA (2024) and V-JEPA 2 (2025) — and DeepMind's Genie (ICML 2024) and Genie 2 (research announcement, Dec 2024) learn interactive environment models from video. This family concentrates on physical scenes, simulated environments, and control; TargetSpace borrows its measurement intuition — preserve predictable, task-relevant structure rather than the full sensory record — and applies it to a persistent individual. The relationship is architectural family versus evaluation framework: a JEPA-style latent predictor is an eligible participant in the architecture-by-evidence grid, not a competitor, and TargetSpace mandates no world model at all.

Live forecasting platforms. Prophet Arena (preprint, 2025) and KalshiBench (preprint, 2025) score LLM forecasts of real-world events resolved after training cutoffs, with calibration front and center — the same sealed, proper-scored, calibration-gated machinery TargetSpace adopts, and strong evidence that prospective forecasting evaluation is maturing. The difference is the object: public external events versus the latent state of one consenting, tracked person. Without a tracked individual there is no own-routine baseline to admit and no wrong-target permutation to run — the two controls that make skill target-specific — which is precisely the layer TargetSpace adds. Both platforms are preprints with live leaderboards, cited as such.

5 · Commercial convergence

The product category is converging on the same promise from different directions, per official pages reviewed July 2026: Limitless (pendant capture; acquired by Meta, December 2025), Bee (“the wearable AI that understands you”; acquired by Amazon, July 2025), Omi (open-source “second brain,” 300,000+ professionals claimed), Plaud (“No.1 AI note-taking brand,” two million professionals claimed), TwinMind (“never forget anything & always know what's next”), and Personal AI (“the AI memory platform,” now carrier infrastructure). None of this is a criticism: these are products, evaluated publicly the way products are — adoption, ratings, time saved, demos, and in two notable cases real quantitative benchmarks of adjacent layers (TwinMind publishes head-to-head transcription accuracy; Personal AI publishes serving latency and token cost). The pattern is the point: the perception and serving layers get numbers; the persistent-model-of-you layer, where the category's central promise lives, is not yet publicly measured. TargetSpace offers a way for any of these companies to test the stronger claim prospectively, federated, without raw data leaving their infrastructure.

6 · What remains unstandardized

Across everything reviewed, no public specification combines: prospective forecasts issued before outcomes exist · external commitment (or explicit self-attested labeling) · strictly proper probabilistic scoring · a population-prior baseline R1 · an admitted own-routine baseline R2 · calibration as a validity gate · matched wrong-target permutation · evidence-tier and modality ablation · disclosed coverage and missingness · observation cost and evidence efficiency · minimum sufficient observation · privacy as a vector rather than a scalar · federated or local execution for sensitive data · explicit null-result validity · one reporting contract with verification levels. TargetSpace is not the first project to model individuals, forecast behavior, or evaluate longitudinal evidence. Its contribution is to bind those traditions into one prospective measurement contract.

7 · Why the conjunction matters

Every missing control reopens a specific failure mode:

No R1 → base-rate success reads as skill.  No R2 → routine replay reads as personal knowledge.  No calibration → overconfident luck survives.  No wrong-target gate → generic pattern-matching passes as intimacy.  No sealing → hindsight and contamination leak in.  No ablation → observation claims go unearned.  No cost disclosure → the incentive tilts toward maximal capture.  No coverage manifest → instrumentation artifacts masquerade as behavior.  No verification levels → self-run claims become incomparable.

Adjacent benchmarks test memory, preference adaptation, action anticipation, clinical markers, or person inference. TargetSpace asks a narrower and more demanding question: did the system acquire calibrated predictive skill about this particular target, beyond population priors and beyond that target's own routine?

8 · The comparison, honestly

ProjectLongitudinal evidencePersistent individualFuture predictionProspective sealingProbabilistic calibrationPopulation baselineOwn-routine baselineWrong-target specificityEvidence attributionObservation-cost reporting
Reality MiningYesYesPartialNot centralNot identified*Not centralNot centralNot identified*Not centralPartial
Digital phenotypingYesYesPartialNot identified*Not identified*Not centralPartialNot identified*Not centralPartial
GLOBEMYesYesNot centralNot centralNot identified*PartialPartialNot identified*PartialPartial
PULSEYesYesPartialNot centralNot identified*PartialPartialNot identified*YesNot identified*
Ego4D forecastingPartialNot centralYesPartialNot identified*Not identified*Not centralNot centralNot centralNot central
FineRobYesYesPartialNot identified*Not identified*Not identified*Not identified*Not identified*PartialNot identified*
KnowMe-BenchYesYesNot centralNot centralNot identified*Not identified*Not identified*PartialPartialNot identified*
LTM benchmarks (LoCoMo, LongMemEval…)YesPartialNot centralNot centralPartialNot identified*Not identified*PartialPartialNot identified*
LifeSim-EvalYesYesNot centralNot centralNot identified*Not identified*Not identified*PartialNot identified*Not identified*
TargetSpace v1.0†YesYesYesYesYesYesYesYesYesYes

* “Not identified in public specification” never means the feature is absent — only that it is not described in the primary sources reviewed. “Not central” means the project explicitly targets something else. These projects were designed for different goals; this matrix compares measurement-contract dimensions only and does not rank scientific quality. Every row maps to sources.json.

† TargetSpace values describe the v1.0 protocol specification — design commitments of a pre-pilot proposal with a synthetic harness, not demonstrated results.

9 · Relationship, not replacement

TargetSpace does not replace these research programs; it supplies a common instrument through which some of their strongest systems could be compared. Sensing programs supply the evidence discipline. Memory benchmarks supply components worth testing. Egocentric perception supplies candidate evidence arms. Products supply the deployed capture the question is about. If a team behind any project on this page believes a row understates its public specification, the correction is welcome — this landscape is maintained against public sources and revised as the field develops.