PilotDeck gives corporate buyers a testable way to examine project-level memory controls as they assess agents for long-running, multi-project work.
ModelBest has co-released PilotDeck 2.0, an open-source system for AI agents—software that carries out multistep tasks—with Tsinghua University’s THUNLP lab, OpenBMB and AI9Stars. The October 15 update centers on a named project container called a WorkSpace, which holds a project’s files, saved context and reusable tools. Pandaily reported the release, and QbitAI separately described its memory and routing functions.
The release makes the stored context of an agent a user-facing control. PilotDeck says users can see memory entries, identify where they came from, then edit, remove or pin them. It also says a cleanup process called Dream Mode can organize memories during idle time and restore an earlier state after a rollback.
For companies testing agents across several assignments, that design makes the handling of retained task information easier to inspect before an agent is used for longer work. PilotDeck is available under the AGPL-3.0 open-source license and supports macOS, Linux and Windows, according to the two reports.
WorkSpaces put a boundary around each agent project
A WorkSpace is PilotDeck’s basic unit of organization. Each project receives its own file system, memory store and skill set, according to both reports. The design aims to keep the material an agent uses for one assignment separated from the information and tools associated with another assignment.
That boundary matters when a team has parallel work underway. A group preparing a data report can preserve the documents, remembered instructions and task-specific tools for that work inside one WorkSpace, while a separate task uses another. The release presents the separation as a way to reduce context spillover between projects; the reports describe the intended product behavior.
Pandaily reported that version 2.0 also adds one-click installation, visual configuration and office-document skills, while QbitAI said the system can continue long-duration jobs in the background. The latter capability, called Always-on, is designed to keep monitoring and execution moving after a user leaves the chat window, then save results locally with a progress summary.
Memory controls make the stored context editable
PilotDeck lets users inspect, edit, delete, pin and roll back memories within project-level WorkSpaces. Its “white-box memory” label refers to a user-visible record of memory generation, extraction, storage and retrieval.
The practical test is whether a user can find a faulty instruction or fact without discarding the work around it. QbitAI reported that the interface allows individual entries to be changed, deleted or fixed in place. That gives evaluators specific controls to test with their own documents and task histories.
Dream Mode is the other part of the memory proposition. The project says it consolidates and deduplicates scattered memories during idle periods, then permits a one-click rollback if the result omits an important detail. A rollback is a return to an earlier saved state. PilotDeck’s release thus ties its memory feature to both user intervention and reversibility.
Smart Routing shifts the test burden to each buyer
PilotDeck also assigns different parts of a workflow to different models. The project says Smart Routing sends complex planning calls to stronger models and routine formatting or polishing work to lighter models. The routing feature controls model selection within a workflow.
In a seven-task comparison cited by Pandaily, the team reported a score of 70.6 at US$3.15 for a Claude Sonnet 4.6 and MiniMax-M2.7 combination, against 69.1 at US$18.36 for Sonnet 4.6 alone. QbitAI also reported a social-media workflow in which a configuration using Claude Opus 4.5 and Sonnet 4.5 cost US$2.83, compared with US$12.58 when it used Opus 4.5 throughout. EastFrontier’s April coverage of Alibaba Cloud’s Qwen3.6-Plus enterprise-agent release provides context for the broader push to match models to agent tasks.
Pandaily says all of those figures are self-reported by the PilotDeck team, and neither report supplies an independent replication or full benchmark protocol. Corporate users can install the code and test the workspace boundary, memory controls and routing choices against their own workflows. A disclosed evaluation method and independent repeat tests would give the reported cost and score comparisons a firmer basis.