Proactive Personal Agents
README · Agents & Applications · Projects & Products · Streaming Proactivity · Benchmark Matrix
This branch covers assistants that use a person's goals, preferences, routines, or ongoing context to decide whether, when, and how to help. Personal refers to the relationship with the user, not the device. Desktop, mobile, wearable, and embodied interfaces can all participate.
The collection's umbrella remains proactive agents. A personal chatbot, persistent memory store, or generic multimodal model is included in this branch only when a documented proactive decision or a clearly labeled supporting role connects it to that scope.
Start with Four Questions
| Question | Representative entry | What to read for |
|---|---|---|
| What need can be inferred before a complete command? | Proactive Agent, PersonalAlign / HIM-Agent | Task anticipation and the distinction between filling omitted preferences and suggesting an unrequested routine. |
| What keeps an assistant active between conversations? | OpenClaw, Proactivity SDK, dot | Heartbeats, persistent goals, events, schedules, and user controls. |
| What changes when the assistant sees the physical world? | Satori, ProMemAssist, Vinci2 / EgoMemo | User-state inference, memory, and decisions to intervene or remain silent. |
| How do we know the help is appropriate for this person? | π-Bench, KnowU-Bench, EgoPro-Bench | Hidden intents, consent/rejection, personalized content, timing, and false interventions. |
Implemented Assistants and Runtimes
| Route | Projects or products | Role and boundary |
|---|---|---|
| Ambient desktop context | MineContext | Context processing and proactive summaries/tips; desktop is this project's interface, not a restriction on the whole branch. |
| Connected everyday services | GAIA, OpenAI dot | Service context, background work, briefings, and suggestions. Project/product documentation establishes advertised behavior, not comparative research performance. |
| Persistent personal-agent execution | OpenClaw, Hermes Agent, memUBot | Runtime activation and recurring workflows; inspect the configured policy that decides whether to notify or act. |
| Reusable activation mechanisms | Proactivity SDK, Letta Code | Durable goals, memory, and schedules can support a personal assistant, but are not a complete personalization evaluation. |
The canonical release, licensing, and checked-date entries live in PROJECTS.md. Memory and sensing components are routed through INFRASTRUCTURE.md.
Research across Devices
| Interface / setting | Methods and systems | Matching evaluation or human evidence |
|---|---|---|
| Desktop and cross-app workflows | Proactive Agent, ContextAgent | ProAgentBench, π-Bench, VibeLifeBench. These are complementary protocols, not necessarily each method's reported benchmark. |
| Mobile | PersonalAlign / HIM-Agent | AndroidIntent in PersonalAlign; FingerTip 20K and KnowU-Bench cover additional suggestion, execution, and interaction questions. |
| Conversation, meetings, and learning | PASK / IntentFlow | LatentNeeds-Bench tests demand detection and silence in speech-transcript sessions; staged user evidence remains preliminary. |
| Passive GUI context and intent recommendation | PIRF / PIRA-Bench | PIRA-Bench evaluates profile-conditioned future-intent recommendation and noise rejection; PIRF is its memory-aware, state-tracking baseline. Recommendation is distinct from executing the suggested task. |
| AR glasses and wearables | Satori, ProMemAssist, WatchGuardian | Task-specific user studies; a watch's user-defined behavior trigger differs from a learned multimodal dialogue policy. |
| Continuous egocentric video | Vinci2 / EgoMemo | EgoServe and EgoPro-Bench; recorded video or simulated profiles do not establish longitudinal wearable deployment. |
| Continual embodied collaboration | PACT | Cross-day ask/act adaptation in controlled embodied scenarios. Also belongs in the robot assistance route. |
Models for Personal Interaction
These are candidate interaction components. Classifying a model here does not assert that it maintains a long-term personal profile or runs on the end device.
| Needed capability | Candidate model / framework | Integration question |
|---|---|---|
| Notice visual events and decide whether to speak | JoyAI-VL-Interaction, MOSS-VL-Realtime | How are the person's mandate, preferences, and silence conditions supplied? |
| Preserve useful evidence and answer at the right time | OneStreamer, StreamReady, StreamOV, Response-G1, EyeWO | Is the behavior internal recording, waiting on a standing query, or initiating new assistance? |
| Continue perceiving while speaking or delegating work | MiniCPM-o 4.5, Gander | How does the interaction model coordinate with memory, tools, authorization, and background results? |
Keep model architecture, resources, and availability in STREAMING.md. Keep personal-workflow evidence here. The same model can also support meetings, gaming, or embodied assistance.
Evaluation Checklist
Report the user's standing mandate; what context and history the assistant can see; who selects the next action; positive interventions and silence negatives; usefulness and timing; clarification/notification burden; consent and rejection handling; and the actual device/inference setup. A successful scheduled reminder is evidence of working automation; stronger claims about discovering an unstated need require a corresponding task and evaluation.