Zero-Model Screen Compilation: 86x Agent Memory Compression at 98% Accuracy
Nossa Iyamu · hf · 2026-08-08
Introduces Activity Frames, a deterministic screen-activity compilation method to solve the issue of computer-use agents paying premium inference costs to re-derive user routines.
This pipeline uses zero models to segment screen captures into typed activity frames containing app, timing, and input data:
- Extreme Compression: Tested on a 51-day corpus (128,756 frames), it compresses a day's raw capture 86x into a prompt-ready block in just 68 ms.
- High Accuracy: Agents reading the block answer daily activity questions with 98.4% accuracy, vastly outperforming LLM summaries (66-80%) and allowing mid-tier models to match frontier ones.
- Cost Evaluation: Measures previously unmeasured agent-cost parameters (Routine Overhead Ratio R and recurrence h), indicating a realistic fleet-wide token ceiling near 8%.
The schema, compiler, and evaluation harness are fully open-source.
More from coding & agent
- Over 50 Verified Agent Skills for Claude Code to Automate Workflows — tom_doerr · 2026-08-08
- Three Commands to Wrap Any LLM into a Local OpenAI-Compatible Endpoint — JeremyCMorgan · 2026-08-08
- Harmonic Rebuilds Scout on Deep Agents, Quadruples User Retention — LangChain · 2026-08-08
- A Practical Guide to SSH Tunnels: Local and Remote Port Forwarding — HankYeomans · 2026-08-08
- Ditch the Terminal: Community Launches Open-Source Grok Build Desktop App — PawelHuryn · 2026-08-08
- Claude Code Defaults to Auto Mode, Catching 89% of Dangerous Commands — DrDatta_AIIMS · 2026-08-08