From our lab to the world
Our team publishes peer-reviewed research in causal AI, graph reasoning, and production simulation, and writes about what we're learning along the way.
Our Research Focus
Institutional Memory
REASONARA is our graph-structured causal memory. It holds over 125M tokens of effective context in production today, at 94% accuracy and 98% fewer tokens than full-context retrieval. We're scaling the same architecture to 1B tokens, enough to hold an enterprise's entire decision history alongside its code.
SWE World Model
Our Software World Model maps a codebase into a six-layer causal graph: what the code does, why it exists, who owns it, and how a change spreads. Today that drives a 13.9-point gain in localization accuracy at 30 to 40% fewer tokens. Next comes formal verification. We simulate how a proposed change will ripple through code, infrastructure, and runtime before it ships.
Agent + Human Enterprise
Today the platform captures decision traces, the reasoning behind every change, and feeds them back to both agents and engineers. Our SkillLens research shows agents reusing that knowledge to lift ALFWorld task success from 45.0% to 51.3%. We're turning this into a shared workspace where people and agents can see the thinking behind each other's work and build on it.
Long Horizon Complex Orchestration
Institutional Memory and the Software World Model let agents hold structured context across long, multi-step tasks. That's what separates steady progress from the brute-force flailing that breaks most agents as a session grows. Our 2027 goal builds on it directly: 100-step orchestration that stays coherent where today's agents fall apart.
Research Publications from Causal Dynamics Lab
Cielara Reasonara
Graph-Structured Causal Memory for Agentic Systems
May 5, 2026
Enterprise AI fails not because models are weak, but because memory infrastructure is broken. Cielara Reasonara is a graph-structured causal memory system that gives AI agents persistent, structured knowledge, including the latent constraints, prior commitments, and causal context that standard retrieval systems miss entirely. Achieving 94% accuracy on enterprise-scale benchmarks with 98% fewer tokens than full-context approaches, Reasonara is the only memory system that explicitly addresses cognitive memory: the class of retrieval problems where the right answer depends on what the agent already knows, not just what the query asks.
RCML
Multimodal Representation Learning Conditioned on Semantic Relations
May 9, 2026
RCML is about representation. CLIP-style models squash a multimodal artifact into one fixed embedding and reuse it no matter what you're asking. RCML conditions the embedding on the relation you're querying, so the same sample looks different in different contexts. On relation-conditioned retrieval that's roughly 31% better Hit@5 than the CLIP backbone, and it holds in zero-shot, fine-tuned, and out-of-domain settings.
SkillLens
Adaptive Multi-Granularity Skill Reuse for Cost-Efficient LLM Agents
May 8, 2026
SkillLens is about reuse. Most systems treat a learned skill as a flat prompt block you either dump in whole or rewrite from scratch. We organize skills into a graph and pull only the pieces that fit. On ALFWorld it beat every skill-reuse method we tested (AutoSkill, EvoSkill, the vanilla baseline), lifting success from 45.00% to 51.31% with 50 steps tasks.
LARGER
Lexically Anchored Repository Graph Exploration and Retrieval
May 8, 2026
LARGER is about navigation. Hand the agent a dependency graph of the whole codebase+context before it opens a single file. GPS instead of driving down every street. We saw +13.9 points on localization, about 10% faster, and 30 to 40% fewer tokens.