Tsinghua's agent OS audits its own memory and haggles over tokens

What it does

PilotDeck is an open-source agent operating system organized around “WorkSpaces”—self-contained project environments that isolate files, memory, and skills. It targets long-running, multi-task productivity rather than one-shot chats, offering a unified runtime across Web, CLI, and instant-messenger front ends with native Model Context Protocol support.

The interesting bit

Memory is fully white-box: every entry is visible, editable, scoped to its WorkSpace, and reversible via a “Dream Mode” that consolidates idle-time state with one-click rollback. The system also auto-routes tasks by difficulty, dispatching grunt work to lighter models and reserving heavyweights for planning checkpoints. The team claims this cuts token costs by roughly 70% on social-media workflows and delivers one-sixth the spend on complex tasks while matching or beating single frontier-model performance.

Key highlights

• Per-project isolation prevents context pollution between parallel WorkSpaces
• End-to-end memory audit trail: generation, extraction, storage, and retrieval are all inspectable and editable
• Smart Routing with self-reported benchmarks showing significant cost reductions against monolithic and all-frontier baselines
• Always-on background execution that discovers tasks, runs monitors, and lands deliverables as local files
• Native MCP support and consistent behavior across Web, CLI, and IM interfaces
• AGPL 3.0, developed by Tsinghua THUNLP, ModelBest, OpenBMB, and AI9Stars

Verdict

Worth a look if you manage multiple long-horizon agent projects and want memory you can debug like a database rather than debug by starting over. Skip it if you’re after a lightweight IDE plugin or a simple chat wrapper.