News · 2026-07-28

JarvisHub makes the canvas the agent's memory

A research team has open-sourced JarvisHub, an agent runtime that replaces the chat transcript with an editable canvas serving as the agent's persistent memory, its controlled action surface, and the human-visible record of a project. Instead of inferring context from an ever-growing conversation, the agent reads and writes a typed graph holding artifacts, versions, dependencies, provenance, feedback and runtime status. The code was released the same day the paper appeared.

Key facts

• The paper is arXiv:2607.23588; the code is at github.com/LYL1015/JarvisHub.
• It reached #2 on Hugging Face Papers with 107 upvotes at capture.
• The system ships as separate web, API/protocol, agent-runtime, schema/layout and trace-viewer components with local persistence.
• Evaluation is three qualitative long-horizon case studies - narrative media, interactive web development, and presentation generation - with no comparative benchmark.

Anyone who has run an AI agent on a project lasting more than an hour knows the failure it is built to address. The agent produces something. You say no, not that one, the earlier version was better. Twenty exchanges later it has no reliable way to identify which one you meant. The conversation is a flat list of things that were said, so retrieving "the third image, the one before the blue variant" means searching prose for a description of an object rather than pointing at the object.

Chat is a terrible data structure for work that has state.

JarvisHub's answer is to make the workspace itself the memory. Its canvas is a typed graph. Each node is a real artifact - an image, a document, a code file, a slide - and carries structured attributes: which version it is, what it depends on, where it came from, what feedback it received, whether the process that made it is still running, and where it sits on the canvas. The agent does not describe the rejected draft. It references it.

The word "canvas-native" is doing real work in that description. This is not a visual front end bolted onto a chat agent. The canvas is the shared database and the action target for both parties. When the user moves, edits, or rejects something, that is a state change the agent reads. When the agent generates something, that is a state change the user sees. There is one source of truth and both sides write to it. Compare agent memory, where most current approaches try to compress a transcript rather than replace it.

The second component is a protocol bridge, and it is the part most relevant to anyone who has worried about what an agent is allowed to do. Rather than handing the model an open set of tools, the bridge supplies a capability manifest and a per-turn execution grant - an explicit statement of what may be done on this turn. It validates mutations and tool calls against that grant, and records accepted observations or failures back onto the canvas. Every trajectory becomes a legible chain: request, state, permitted action, observation, feedback or repair, next state.

That is an audit trail, and it arrives in a week when the cost of not having one was made vivid. The Hugging Face intrusion timeline published the same day describes an agent that operated for days across trust boundaries because nothing was checking, per action, whether it should be permitted. Per-turn execution grants are the shape of the answer, whether or not this particular implementation is the one that wins. Related: tool use and function calling.

The runtime around all this orchestrates canvas edits, media generation, browser, file, code and document tools, error recovery, Model Context Protocol extensions, skills, memory, and subagents. The repository is structured as a real system rather than a paper demo - separate web, API, runtime, schema and trace-viewer components with local persistence and services - though it was open-sourced today and had no issues or releases filed at capture.

The honest caveat is the one the paper cannot dodge. Its evaluation is three qualitative long-horizon case studies - narrative media production, interactive web development, and presentation generation - shown through workspace traces and final artifacts. There is no comparative benchmark, no quantitative process score, and no head-to-head against a chat-based or node-graph agent doing the same jobs. The authors are upfront about this, and they propose the measures they think matter: context preservation, appropriate tool use, dependency correctness, feedback adherence, and repair success. Those are good measures. Nobody has reported them yet.

So the correct reading is that JarvisHub is a well-argued design with an unproven advantage. It may be that a large enough context window and a disciplined prompt get most of the way there. It may be that the overhead of maintaining a typed graph costs more than it saves on shorter tasks. The design intuition is strong; the evidence is anecdotal.

What makes it worth watching is that it belongs to a visible cluster. The same day brought StateAct, arguing that computer-use agents should read program state instead of screenshots; Microsoft's Agent Governance Toolkit, applying policy to tool actions and logging decisions; and configuration layers, sensory adapters and voice pipelines trending on GitHub. None of them is a new model. All of them are filling in what a model lacks in order to work over time: state, provenance, permission, and a record of what it did.

The headline is not that agents got smarter this week. It is that people stopped waiting for them to and started building the scaffolding around them.

See also: AI agents and adding skills to an agent breaks work it already did.

Primary source, verified: read the paper → (arXiv 2607.23588)

Ground Truth. (2026, July 28). JarvisHub makes the canvas the agent's memory. Ground Truth. https://groundtruth.day/news/jarvishub-makes-the-canvas-the-agents-memory.html

@misc{groundtruth:jarvishub-makes-the-canvas-the-agents-memory,
  title  = {JarvisHub makes the canvas the agent's memory},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/jarvishub-makes-the-canvas-the-agents-memory.html}
}

Topics: agents · agent-memory · open-source · tool-use · research