Research

Our research focuses on making AI agents safe, controllable, and useful in real-world environments.

AI Safety & Control

Human-in-the-loop approval systems, safety gates, kill switches, and behavioral boundaries for autonomous agents. How do we keep AI controllable as capabilities scale?

Internal research ongoing

Agent Memory Architecture

Hybrid semantic + episodic + procedural memory with HNSW vector search, importance-based consolidation, and decay functions. Building agents that remember like humans do.

Memory Engine v2 (2026)

Agent-to-Agent Communication

Protocols for inter-agent task delegation, capability discovery, and trust establishment. How do autonomous agents collaborate safely?

A2A Protocol Spec (2026)

Emotion & Personality Modeling

LLM-native emotion systems with 20-dimensional affect space. No separate models, no GPU requirements. Personality as a first-class runtime parameter.

Persona Shell Architecture (2026)

Real-Time AI Presence

Sub-millisecond state synchronization for AI agents in social environments. Presence, attention, and natural turn-taking in voice conversations.

Research phase

Evaluation & Self-Improvement

Automated evaluation of agent performance with RLHF-lite feedback loops. Accuracy, honesty, execution quality, and discipline metrics with hidden MMR ranking.

Rank System v1 (2026)