See what your
agent forgets

Same AI. Same game. One remembers.

We take any AI model — even the cheapest one — and put it in a game where every decision is irreversible, every mistake accumulates, and the clock is ticking. Then we run it twice in parallel: once with Bridge (a 15-line mechanism that lets the AI leave notes to itself), once without.

Bridge ON (left) stays near the goal. Bridge OFF (right) drifts into chaos. Same model, same game, same moment.

The W trajectory tells the whole story. Teal = Bridge ON. Coral = Bridge OFF. The divergence is irreversible.

This is not a better prompt. Not a bigger model. Not RAG or external memory. It's 15 lines of code that let the AI leave itself a note between turns. That's the entire mechanism.

Validated across 7 architectures

The counterintuitive finding: Bridge works BETTER on small, cheap models than on large ones. A stabilized Haiku outperforms an unstabilized model 10× its size beyond the complexity threshold. This changes the economics of AI agents in production.

OMEGA-TRACE: irreversible decisions under constraint

You move on a hexagonal grid. Every edge you cross becomes blue (progress) or grey (more work) depending on your direction relative to the curve. Your goal: complete all edges (W = 0). Simple rules, hard consequences.

Blue edges = completed targets. White/red edges = remaining work. The AI sees the same map a human player would see.

Three constraints make it brutal: you can never reverse direction, you can never repeat the same 5-move sequence twice (K-memory), and a periodic BITE erodes your completed work. The longer you play, the fewer options remain. Oxygen counts down.

Now replace the game with any task: a long conversation, a coding agent, a project manager. The structure is identical — irreversible decisions, accumulating constraints, a goal that drifts further away if you lose focus. Every AI agent in production faces this. None of them know it.

5 minutes. Your model. Your prompt.

Bridge Lab is an open experiment. You bring your own AI, your own API key (it never leaves your machine), and your own prompt strategy. The server runs the game and the Bridge. You watch.

The runner in action. [ON] and [OFF] alternate moves. Your API key stays on your machine — only arrows and notes go to the server.

Download Runner

Not a hack. A structural discovery.

Bridge Lab is the visible tip of a larger research program called Sub-Limit Dynamics (SLD) — a mathematical framework that explains why systems operating under finite constraint all exhibit the same structural properties, from prime number gaps to wildfire spread to AI agent collapse.

The Bridge mechanism is the first patented intervention derived from this framework. It works because the theory is correct: drift is not a bug in the model, it's a structural property of any system making sequential irreversible decisions in a bounded context.

• Sub-Limit Dynamics — Boundary interaction as the structural determinant in finitely constrained generative systems (C3)
• Stateless Agents in Exhaustion Systems — Memory, criterion, and the geometry of failure across 7 LLM architectures (O3)
• Compensatory Recovery — Regime indicator in bounded-context agents, 170 level-episodes, 95.9% classification accuracy (O4)
• Cross-Domain Isomorphism — Same structure across wildfire, traffic, chemical oscillators, and convection (C8)
• Grammatical Structure of Prime Gaps — SDL validated on the cleanest mathematical domain (C9)
• + 8 more papers covering positional arithmetic, constrained walks, semantic games, music, and spiral dynamics

All papers on Zenodo. Patent: USPTO Provisional 63/993,764, filed March 1, 2026.