A research company building cognitive infrastructure for AI agents.

We research active memory, continual learning, and grounded decision-making — and build the tools that embody what we discover. Model-agnostic. Mathematically grounded. Running in production.

What We Believe

Every problem is soluble.

We're optimists in the Deutschian sense — not “things will be fine” but “every problem has a solution if you look hard enough.” That belief drives everything we build.

The Problems We See

Agents don’t remember

Current AI agents start from zero every session. Memory solutions today are static markdown (.md) files with logical heuristics. There’s no real learning, no transfer across projects, no grounded recall. You’re constantly re-explaining what you’ve already built.

Specifications always drift

Everyone’s talking about spec-driven development, and most people who’ve tried it say it doesn’t work. They’re right — it doesn’t work the way they’re doing it. The bottleneck is a 3,000-line markdown file that no one can keep synchronized with reality.

Your tools work against you, not with you

You’re switching between a CLI, a desktop editor, a browser, and a chat window — copy-pasting between them, losing context every time. Moving to yet another tool feels like more friction, not less. What if everything just worked together, in one place, on any device?

01Foundation

Subtext

Your LLM reasons. Subtext decides what it reasons about.

A cognitive architecture that sits alongside any LLM. It builds active memory from every interaction, matches past experience to current situations, and learns continuously — across sessions, across projects, without fine-tuning.

Active memory that evolves with every interaction — not static markdown (.md) files
Matches past experience to current situations through nonparametric context discovery
Transfers knowledge across projects automatically — nothing starts from zero
Grounded decision-making that reduces hallucinations by anchoring on what actually happened
Context

"The bug might be upstream in the OR handler — let me trace through the evaluation logic..."

Recalled

OR handler fix was already tried and reverted in BM55 — the root cause is in expression-eval, not the handler

Agent Feedback

"Wait — this was already tried. The memory just saved me from going down a dead-end path."

Knowledge
Behavioral
2,599 memories · 3 projects
02Workflow

Blueprint

The bottleneck isn't your agent. It's your specification.

Point it at a codebase — get a living specification. Modify it through a visual interface, not a 3,000-line markdown file. Blueprint generates precise implementation context for the agents you define. Your spec never drifts from your codebase.

Visual specification editor — modify through a GUI, not raw text
Generates precise implementation context for agents you customize
Source-derived specifications with explicit behavioral invariants
Specifications that are always synchronized with the actual codebase
Specification Compilation
Specification
Codebase
App
app.tsx
Auth Flow
auth.ts
Dashboard
Dashboard.tsx
Login
login.tsx
Settings
settings.tsx
API Layer
api-client.ts
Compile / FeedbackDiff generationImplementation
03Interface

Maestro

Everything you need. One window. Any device.

A fully customizable development environment in the browser. You and your agents can pull up files, videos, wireframes, and simulators — all in the same window, positioned exactly how you want them. Never lose a conversation. Never restart because of memory issues. Start on your computer, continue on your phone.

Files, videos, wireframes, and simulators — all in one browser window
Arrange every panel exactly how you want — collapse, resize, reorder
Never lose a conversation or restart because of memory issues
Start on your computer, continue on your phone — always synced

Same workspace. Any screen.

Panels on desktop. Conversation media on mobile.

Maestro keeps chats, files, images, videos, and agent context together. Arrange the desktop exactly how you work, then pick up the same thread from your phone without losing the media your agents surfaced.

Desktop panelsMobile chatSynced media

Technical Foundation

Built on mathematics, not vibes.

Subtext isn't a retrieval layer bolted onto an LLM. It's a full cognitive architecture grounded in active inference and Bayesian mechanics — the same mathematical framework that models biological cognition. Every decision is derived from first principles, not heuristics.

What This Means In Practice

Active Inference

Agents don't just respond — they anticipate. The system minimizes surprise by predicting what you'll need before you ask.

Nonparametric Learning

No predefined categories. The system discovers its own structure from observations, growing more nuanced with every interaction.

Continual Learning

Memory isn't static snapshots. It evolves online — refining, merging, and pruning autonomously as understanding deepens.

Cross-Project Transfer

Knowledge earned in one project transfers to others. Patterns, decisions, and insights compound across your entire workflow.

Intelligent Cost Reduction

A Bayesian context controller that learns what to keep and what to remove from the agent's context — reducing token spend while preserving what matters.

Model Agnostic

Works with any LLM — Claude, GPT, Gemini, Llama, or whatever comes next. The harness doesn't care what model you use.

Why this is different.

The harness problem.

In 2025, Meta paid $2–3B to acquire Manus — not for a model, not for a dataset, but for an agentic harness. The orchestration layer that makes LLMs actually finish tasks.

Subtext is the next generation. Where current harnesses react, Subtext predicts. Where they forget, Subtext learns. Where they start from zero, Subtext transfers. Built on active inference and Bayesian mechanics — not prompt chains and heuristics.

The specification problem.

Everyone's talking about “spec-driven development” — and most of them have raised their hand to say it doesn't work. They're right. It doesn't work the way they're doing it.

The bottleneck has never been the agent. It's the specification. You spend four hours planning so the AI can finish in one — and you're still re-reading a 3,000-line markdown file to make sure nothing's missing. That's not sustainable. That's not engineering.

Blueprint makes spec-driven development actually possible. It extracts a living specification from your codebase, gives you a visual interface to modify it, and generates the precise diffs that drive your agent pipelines. Your spec never drifts. Your agent never hallucinates a requirement. You get out of the way and let it converge.

Last Generation

Reactive orchestration

Heuristic routing · Stateless · Starts from zero

Current "Memory"

Vector stores + RAG

Cosine similarity · No real learning · Static snapshots

Subtext + Blueprint

Predictive cognition + living specifications

Active inference · Learns · Transfers · Never drifts

How We Work

Conjecture. Criticism. Iteration.

This isn't a tagline — it's our engineering loop. A specification is conjecture. Testing it against the product is criticism. Revision is iteration. Every product starts as a conjecture about how something should work, gets tested against reality, and improves. We don't ship features. We accumulate knowledge.

“All problems are caused by a lack of knowledge, and this knowledge can be attained.”

David Deutsch

Build With Us

We're onboarding builders who want to work at the frontier of agent cognition.

Request free beta access below, or subscribe to Maestro for $10/month. Connect your computer and use your own Codex CLI or Claude Code CLI account; model credits are not included.

Get Maestro · $10/month

Or sign in to manage your account.