Research

What we're working out.

We investigate how AI becomes durable, efficient, and capable of operating inside real systems — work that sits at the intersection of cognitive science, software architecture, knowledge representation, and applied AI engineering.

Most AI research optimizes the moment of response. We're more interested in what happens around it: how a system pursues a goal across time, how it remembers and revises what it learns, when its reasoning should stop being probabilistic altogether, and whether principles from biological cognition can buy adaptability at lower computational cost.

Below are the briefs from our active research program, most recent first. Each reflects a longer-term research direction — and a practical lesson from building AI-enabled systems for complex operational environments.

Jeremy Sublett Co-Founder

Hybrid Graph-Vector Memory Systems

Ontology Induction & Machine Concept Formation

Vector retrieval answers one question well: what resembles this? But two passages can share language while describing different people, different periods, or contradictory claims — and two others can share no language at all while being linked by an entity, an event, or a cause. Similarity alone is not memory.

We're researching memory that layers four capabilities: vectors for associative recall, graphs for identity and relationships, ontologies for valid concepts and constraints, and locally operated models for private, repeatable extraction. The open question is whether a system can move from retrieving passages to forming concepts — separating what is stable from what is merely recurrent, and revising its beliefs as evidence arrives.

  • associative and relational memory
  • ontology induction
  • concept formation
  • provenance
  • belief revision
  • local AI
  • knowledge consolidation
Read the hybrid memory brief →
Jeremy Sublett Co-Founder

SmartFlow

Cognitive Execution & Durable AI Workflows

Most AI systems are built to produce an immediate response. Real work unfolds through sequences of actions, decisions, interruptions, approvals, failures, and recoveries. SmartFlow examines how AI-directed work can persist across time while staying observable, governed, and recoverable.

The central question is where to draw the boundary between probabilistic reasoning and deterministic execution. A model can interpret an objective, select a capability, or propose a next action — but durable execution still demands explicit state, controlled transitions, validation, and recovery. The aim is long-running work that never surrenders operational control to an opaque model.

  • durable execution
  • cognitive control
  • stateful workflows
  • human oversight
  • recovery
  • evaluation
  • governance
Read the SmartFlow brief →
Jeremy Sublett Co-Founder

Hard Logic

Progressive Compilation of AI Reasoning

Language models are valuable when business rules are buried in policies, specifications, examples, and human explanation. But once a decision has been sufficiently understood, asking a model to make that same call again and again adds cost, latency, variability, and operational risk.

Hard Logic studies a progressive path. The model interprets ambiguous requirements, proposes decision rules, and surfaces exceptions; those decisions are then observed, tested, and gradually converted into decision tables, validation functions, and other deterministic mechanisms. Routine reasoning moves out of inference and into repeatable, auditable execution, while the model stays available for the novel and the genuinely ambiguous. The objective is not to eliminate AI, but to use it to discover and engineer durable software logic.

  • rule induction
  • semantic parsing
  • program synthesis
  • neuro-symbolic reasoning
  • decision modeling
  • progressive formalization
  • reasoning compilation
Read the Hard Logic brief →
Jeremy Sublett Co-Founder

Biologically Inspired Cognitive Architectures

More Useful Cognition, Less Computation

The brain supports perception, memory, learning, and adaptive behavior on roughly the power of a household light bulb. It isn't a fair benchmark — different substrate, different tasks — but it is strong evidence that useful cognition doesn't require continuous, uniform computation at enormous scale.

This work isolates specific mechanisms — sparse activation, predictive processing, gated working memory, complementary fast and slow learning — and states each as a testable computational hypothesis measured against a conventional baseline. Current lines of inquiry include whether neurons accomplish credit assignment without a global backward pass, whether neural activity is better modeled as a queue of discrete events, and how much the three-dimensional structure of biological systems contributes to their efficiency.

  • sparse computation
  • predictive processing
  • credit assignment
  • event-driven simulation
  • spatial computation
  • memory gating
  • continual learning
  • fast and slow learning
  • specialization
  • energy efficiency
Read the cognitive architectures brief →

How this connects to client work

These are research directions, not products. But they come directly from problems we hit building production systems in regulated environments — workflows that had to survive interruption, memory that had to preserve provenance, decision logic buried in policy documents, and inference budgets that had to stay bounded. What we learn here informs what we ship, and what we ship keeps the research honest.

Want to talk about this work?

Start a conversation See what we've shipped