Robologue
Built at The Harness Engineering & Model Wrangling Hackathon · Sep 26, 2026 · New York, NY
Robologue: a dialogue for robots. Procedural memory for long-horizon tasks. Long-horizon robots fail for a simple reason: they forget what went wrong five minutes ago. Manufacturing has a rule for this. The 1-10-100 rule says a defect caught during assembly costs $1 to fix and $100 once it ships. A robot with no memory ships every defect twice. Robologue remembers. Unresolved issues stay open across interruptions, restarts, and unrelated work until an explicit resolution event clears them. Replays are idempotent by SHA-256 digest. Every decision persists per recording in MongoDB Atlas, and the harness reads that history back across restarts. Memory that survives the process dying. Deployed robots don't learn. Retraining on new data costs human time and money, so most systems ship frozen and stay frozen. Robologue improves while it works: the gate turns the robot's own mistakes into better checks, no retraining run required. A budgeted vision-model client extends the same checks to real assembly frames, with hard caps on model calls and spend.