AI app-market evidence brief

Scope and staleness

• Evidence below is limited to stored notes cited here, mostly collected between 2026-06-26 and 2026-08-12. That is recent relative to 2026-08-17, but the corpus is thin for dedicated smart-glasses hardware launches, consumer-spatial platforms, and first-party market-adoption data.

Facts from stored notes

1. The stored notes show repeated movement toward on-device and locally deployable agents, not just cloud assistants. PalmClaw describes a native mobile-phone agent framework that exposes device capabilities as explicit tools and reports an 11.5% relative task-success improvement plus a 94.9% completion-time reduction versus its strongest baseline. SelectInfer describes neuron-level selective loading and computation for edge LLM inference while preserving task performance. A separate benchmark on longitudinal data-preparation tasks reports that open-weight 31B-35B models running on consumer-grade hardware reached up to 87.9% average task completion in governance-restricted settings. PalmClaw: A Native On-Device Agent Framework for Mobile Phones (http://arxiv.org/abs/2607.13027v1), SelectInfer: Selective Neuron Loading and Computation for On-Device LLMs (http://arxiv.org/abs/2607.18081v1), Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks (http://arxiv.org/abs/2607.21482v1)

2. Several notes point to persistent personal-memory systems as a core capability layer for consumer and wearable apps. LightMem-Ego describes continuous capture of egocentric visual and audio streams into current, short-term, and long-term memory, with deployment on smartphones and AI glasses for object finding, conversation recall, life summarization, routine discovery, and personalized assistance. Vinci2 frames proactive assistance in continuous egocentric video as a decision problem about when help is welcome, and its EgoServe benchmark spans more than 3,000 service instances across 10 categories and four memory horizons. R4DSG adds a structured memory layer for wearable assistants and AR systems by preserving persistent object identity and anchor-relative spatial change, reporting a 6.7-point gain over EgoRAG-Text on an object-related EgoLifeQA subset and a 12.5-point gain on when questions. LightMem-Ego: Your AI Memory for Everyday Life (http://arxiv.org/abs/2607.11487v1), Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos (http://arxiv.org/abs/2607.11523v1), R4DSG: Relative 4D Scene Graph Memory for Object-Centric Question Answering in Long Egocentric Video (http://arxiv.org/abs/2608.11017v1)

3. In the vault, voice is appearing as a serious interface layer, but the evidence also says it remains operationally weak. DuplexWorld evaluates voice agents across banking, insurance, travel, healthcare, logistics, and pathfinding, and reports that even the best tested systems still left substantial room for improvement on task success, turn-taking, and speech naturalness. A needs-review note on HIVE reports that voice-transcription perturbations lowered accuracy across all tested instruction-tuned models, that the damage came mainly from altered transcription structure rather than fillers, and that extra reasoning budget largely recovered keyboard noise but not spoken-register degradation. DuplexWorld: Can voice agents help you get through the day? (http://arxiv.org/abs/2608.10716v1), 2026-08-04-should-we-type-or-talk-to-llm-agents-a-comprehensive-study-of-voice-a (http://arxiv.org/abs/2608.03970v1)

4. The notes suggest that agent patterns are shifting from raw GUI imitation toward explicit tool use, orchestration, memory, and policy layers. PalmClaw exposes native device capabilities as tools rather than long GUI action chains. The MCP architecture paper catalogs recurring server patterns such as Resource Gateway, Tool Orchestrator, Stateful Session Server, Proxy Aggregator, and Domain-Specific Adapter, and notes declining tool-selection accuracy as tool counts rise. The GUI-agent planning note reports that high-level task training and hindsight-generated training data improved out-of-distribution planning for small multimodal GUI agents, including a 7B model exceeding a larger baseline on real-world benchmarks. PalmClaw: A Native On-Device Agent Framework for Mobile Phones (http://arxiv.org/abs/2607.13027v1), MCP Server Architecture Patterns for LLM-Integrated Applications (http://arxiv.org/abs/2606.30317v1), Empowering GUI Agents via Autonomous Experience Exploration and Hindsight Experience Utilization for Task Planning (http://arxiv.org/abs/2606.27330v1)

5. The stored notes also point toward consent-aware, longitudinal, human-state-centered agents rather than one-shot task bots. SovereignPA-Bench evaluates user-owned personal agents on privacy, consent, evidence, burden, manipulation resistance, and auditability, not only task completion. ComBodied Agents proposes a closed loop of multimodal perception, longitudinal correctable memory, personal world models, and admissible intervention policies using software tools, sensors, wearables, robots, and human services as action channels. SovereignPA-Bench: Evaluating User-Owned Personal Agents under Evolving Intent, Platform Mediation, and Consent Constraints (http://arxiv.org/abs/2607.05363v1), ComBodied Agents: a New Paradigm of Human-Centric Agentic AI (http://arxiv.org/abs/2608.10915v1)

6. The vertical examples in the vault are concentrated in privacy-sensitive or coordination-heavy domains rather than broad mass-market chat. AdaHome describes a local small-model smart-home assistant with intent-aware planning, up to 3x lower latency, and 88% preference consistency in multi-turn scenarios. ECHO describes a locally deployable chronic-care assistant with persistent memory, 17 clinical tools, a safety layer, and multimodal speech assessment, running on consumer hardware with no patient data sent externally. AI Tour Meeting uses multiple persona-based LLM agents to coordinate group travel itineraries under differing constraints and preferences. AdaHome: An Adaptive Smart Home Assistant using Local Small Language Models (http://arxiv.org/abs/2607.18034v1), ECHO: A Locally-Deployable Agentic Health Assistant with Temporal Memory, Safety Guardrails, and Speech Assessment (http://arxiv.org/abs/2608.06110v1), AI Tour Meeting: Group Travel Planning by LLM Agents (http://arxiv.org/abs/2607.18806v1)

Interpretation for app-market work

1. Interpretation: the most supported greenfield pattern in this corpus is not a generic chatbot, but a local or edge-resident assistant with persistent memory, explicit tools, and bounded intervention policies for privacy-sensitive contexts such as phone-native workflows, smart home control, chronic care, and governance-restricted knowledge work. This interpretation is grounded by the combined direction of PalmClaw, SelectInfer, AdaHome, ECHO, and the local open-weight agent benchmark. PalmClaw: A Native On-Device Agent Framework for Mobile Phones (http://arxiv.org/abs/2607.13027v1), SelectInfer: Selective Neuron Loading and Computation for On-Device LLMs (http://arxiv.org/abs/2607.18081v1), AdaHome: An Adaptive Smart Home Assistant using Local Small Language Models (http://arxiv.org/abs/2607.18034v1), ECHO: A Locally-Deployable Agentic Health Assistant with Temporal Memory, Safety Guardrails, and Speech Assessment (http://arxiv.org/abs/2608.06110v1), Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks (http://arxiv.org/abs/2607.21482v1)

2. Interpretation: the clearest wearable and spatial-adjacent opportunity in the current vault is memory and retrieval over egocentric streams, especially object memory, routine recall, and proactive assistance, rather than fully realized spatial-computing platforms. The note set supports smartphone-plus-glasses assistance layers more directly than it supports immersive XR ecosystems. LightMem-Ego: Your AI Memory for Everyday Life (http://arxiv.org/abs/2607.11487v1), Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos (http://arxiv.org/abs/2607.11523v1), R4DSG: Relative 4D Scene Graph Memory for Object-Centric Question Answering in Long Egocentric Video (http://arxiv.org/abs/2608.11017v1)

3. Interpretation: voice appears better supported as a workflow surface for narrow, tool-backed verticals than as a frictionless general interface. The evidence base here points to opportunity in domains where turn structure, verification, and fallback can be tightly designed, but it does not support a broad claim that voice agents are already mature for general-purpose consumer use. DuplexWorld: Can voice agents help you get through the day? (http://arxiv.org/abs/2608.10716v1), 2026-08-04-should-we-type-or-talk-to-llm-agents-a-comprehensive-study-of-voice-a (http://arxiv.org/abs/2608.03970v1)

4. Interpretation: the emerging moat in consumer and vertical agents may sit less in the base model and more in scaffolding around memory, tool boundaries, orchestration patterns, consent policy, and auditability. That interpretation follows from the MCP pattern paper, GUI-agent planning work, SovereignPA-Bench, and ComBodied Agents. MCP Server Architecture Patterns for LLM-Integrated Applications (http://arxiv.org/abs/2606.30317v1), Empowering GUI Agents via Autonomous Experience Exploration and Hindsight Experience Utilization for Task Planning (http://arxiv.org/abs/2606.27330v1), SovereignPA-Bench: Evaluating User-Owned Personal Agents under Evolving Intent, Platform Mediation, and Consent Constraints (http://arxiv.org/abs/2607.05363v1), ComBodied Agents: a New Paradigm of Human-Centric Agentic AI (http://arxiv.org/abs/2608.10915v1)

Plain gaps in the current corpus

1. The corpus is insufficient to support strong claims about current smart-glasses hardware vendors, shipment/adoption levels, platform economics, or first-party spatial-computing roadmaps. The stored evidence is mostly about wearable-assistant memory and egocentric perception, not hardware market structure.

2. The corpus is insufficient to support strong claims about underserved consumer categories by revenue, install base, or retention. The notes are research-heavy and capability-heavy, with little direct market-usage evidence.

3. One cited voice-input note is marked needs_human_review: true, so it is useful directional evidence but weaker than the canonical source-note set. 2026-08-04-should-we-type-or-talk-to-llm-agents-a-comprehensive-study-of-voice-a (http://arxiv.org/abs/2608.03970v1)