Isaac is an EEG-adaptive conversation agent built on the Arctop platform. It holds an ordinary open-ended conversation -- casual chat, teaching, brainstorming, focused work -- while a real-time signal of the user's cognitive state, decoded from a consumer EEG headset, conditions how the agent communicates: its pacing, complexity, tone, and, when necessary, whether it stops mid-response and changes course.
Isaac is live in production as a first-party tab in the Arctop app.
Dual-agent architecture. Isaac runs two cooperating LLM agents:
- The conversation agent is the user-facing voice. It converses normally and never sees raw EEG.
- The cognitive analyst (recommendation agent) works behind the scenes. It reads the EEG-derived workload signal, maintains a shared memory document of observations about the user and which adaptation strategies measurably worked, and hands the conversation agent short recommendations on what to adjust -- never what words to say.
Sentence-level streaming with cognitive updates between sentences. Responses are delivered sentence by sentence, as text and as speech. Cognitive-state updates arrive continuously between sentences. When a sustained workload shift is detected mid-response, the analyst is consulted while the stream continues; if it judges that continuing unchanged would keep the user overloaded, the response is interrupted and restarted with the adjustment folded in. Interventions are then scored against the measured workload change they produced, and that evidence feeds back into future recommendations.
See ARCHITECTURE.md for the closed-loop architecture, prompts/ for the current production prompts of both agents, PROVENANCE.md for how every session is attributed to the exact prompt build that produced it, and config-schema.md for the deployment configuration surface.
The deployed system measures and adapts on one EEG-derived metric: cognitive workload (a 0-100 score with per-session baseline and trend). Multi-metric cognitive decoding -- enjoyment, focus, flow, stress -- is part of the platform design, not the shipped measurement set. Every adaptive behavior in production Isaac is driven by workload alone, and the analyst is explicitly instructed never to infer other metrics from it.
Isaac is the companion platform described in Section 4 of the preprint Reinforcement Learning from Brain Feedback (RLbF): the deployed instantiation of the paper's closed-loop architecture, in which an EEG-derived cognitive state signal conditions language generation and measured cognitive outcomes score each adaptation.
Preprint: Kay, E.; Furman, D.; Kogan, B.; Chiang, K.-J. Reinforcement Learning from Brain Feedback (RLbF) for Large Language Model (LLM) Improvement: Using and Evaluating Real-Time Neurophysiological Reward Signals for System Adaptation. Preprints 2026, 2026081274. https://doi.org/10.20944/preprints202608.1274.v1
For a shorter, plain-language read on the concepts, see the companion blog post: Teaching AI to Read the Room.
Isaac ships as the Isaac tab inside the Arctop app, paired with an Arctop-supported EEG headset:
- Android (Google Play): https://play.google.com/store/apps/details?id=com.arctop.android
- iOS (Apple App Store): https://apps.apple.com/us/app/arctop/id6787000987
The app also includes a post-hoc visualization (the Cognitive Attention Map) that renders a completed session as a timeline aligning the conversation transcript with the workload curve, detected shifts, analyst recommendations, and interruptions.
| File | Contents |
|---|---|
| ARCHITECTURE.md | The seven-component closed loop and the dual-agent realization (with architecture.png) |
| prompts/ | Current production prompts: conversation agent (version 2) and cognitive analyst (v2) |
| PROVENANCE.md | How session recordings are stamped with model identity and prompt version + fingerprint |
| config-schema.md | Deployment configuration: structure and semantics |
| LICENSE | CC BY 4.0 |
The documentation and figures in this repository are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).