What are claims?
A claim is a durable fact or conclusion tied to your workspace:- Content: The actual statement or finding
- Sources: Evidence supporting the claim (code cells, files, other claims)
- Type: How the claim was created
- Recall priority: How readily agents should see it again
Types of claims
Evidence claims come from three capture paths — notebook capture, script capture, and file capture. Inspect evidence through a claim’s sources or a direct citation.Runtime-captured
Captured from a notebook or script when code produces a value. Required task variables are captured automatically in notebooks; scripts record values through the qualia package. Each claim links to the cell or script that produced it.File-captured
A quote from a workspace file, with its path and location. Qualia checks the quote against its source and flags it if the text changes or the file becomes unavailable. The original quote remains as a record of what was captured. You can create these claims yourself with Save as evidence.When a cited file changes
The Knowledge list flags quotes that no longer match their source or whose source cannot be read. Inspect the file and ask Qualia to update the evidence if needed. Agents using that evidence are also notified of the change. A claim remembers where its notebook was when it was captured. Renaming or moving a notebook in Qualia usually keeps its claims pointing at it. If a claim can’t find its notebook, for example because it was moved while Qualia was closed, the claim shows the notebook as unavailable until you open the notebook again.Figures
A captured plot links back to its notebook cell or producing script. Supported figures include the chart values, so you can inspect points and toggle series. Use save_figure to retain chart data when saving a figure from a script.LiveData and live figures
Ask the agent for a live figure to follow training, solver, or simulation output. Figures update as observations arrive, catch up after reconnecting, and preserve zoom and hidden series. Independent curves keep their own points and gaps. For manual logging from a Qualia-launched Python notebook or script:flush() waits for committed data; to_pandas() returns a detached snapshot.
Failed deliveries can be retried with flush() while the local journal survives.
Static exports stay pinned; sharing future observations
requires explicit approval and does not upload local-only data.
For existing W&B logs, ask the agent to follow a local run directory or its parent
wandb/ directory and choose the metrics you want to plot. Keep the workspace
runtime running to receive updates. SDKs 0.19.11 and 0.30.0 are supported; remote
logs are not supported. Updates may lag while W&B buffers observations.
Synthesized
Written by agents to summarize understanding:- Combine multiple sources (cells, files, other claims)
- Represent conclusions drawn from evidence
When an agent works on a task, it often creates a synthesized claim before
marking the task complete. This provides a record of what we learned by
working on this task.
Related claims
Related claims connect findings on the same subject without treating one as evidence for the other. Find them in the Related section of a claim’s card; click one to inspect it and its sources.Bringing existing work into knowledge
Work that predates Qualia, such as old notebooks, scripts, results files, and papers you read, can be turned into claims in one pass. On the new chat screen, choose Import knowledge and add whatever the work is: a file or folder already in your workspace, files dragged in from your machine, a link, or a note you type. Each one is staged, and starting the import hands them all to a single agent that reads them and records what they establish. Imported sources are saved in the workspace’ssources/ folder so claims can link back to them. PDFs retain page references; web pages are saved as text.
Use files with readable text. Scanned PDFs, images, and archives cannot be imported this way. For pages behind a login or subscription, download a readable copy and upload it yourself.
Recall priority
Every claim carries a recall priority — the agent’s judgement of how readily future agents should see it again. It describes the claim’s standing, not how confident anyone is that it’s true, and not how relevant it is to whatever an agent happens to be doing right now. It applies to all knowledge, so a captured value can outrank a written finding.
Most knowledge is Low, and that’s the intended default — it means “recall
this when it matters”, not “this doesn’t matter”. Reserve High for canonical
facts and constraints that every agent in the workspace should receive.
High claims are injected into the AI’s context automatically, so a
finding an agent files or revises reaches the other agents at their next step
without anyone re-explaining it. You can sort the knowledge list in Project
view by priority.
The knowledge list in Project view
Click Project in the top right. Knowledge appears beside the task map, or below it in a narrow pane.Search and sorting
- The list includes findings, tables, and figures, including live figures; underlying LiveData sources are not listed
- Search: Find claims by content or ID
- Sort: Newest/oldest, or by recall priority
Viewing claims
Each claim card shows:- Claim content (truncated for long claims)
- Type indicator
- Source count
Browsing knowledge in Project view
Click a claim to see its full content and sources. Follow citations and related claims, then use the breadcrumb trail to return to the list. Clicking a citation in chat opens Project and reveals the referenced claim.Working with claims
Saving a selection as evidence
Agents aren’t the only ones who can file evidence. Select any text in a file or in a notebook cell’s text output, and a Save as evidence button appears next to the end of the selection once you release it. It’s also available as Save Selection as Evidence in the menus and the command palette. Selecting inside a cell’s code doesn’t offer it — a claim made from what that code produced already links back to the cell. Neither does text Qualia highlights for you, such as the cited lines a claim’s source opens to or a match in the find bar; select it yourself to quote it. To quote a PDF, switch it to the Text view using the toggle above the document. That shows the PDF’s text layer — the same text an agent reads — and selecting in it works like selecting in any other file. The claim records the page the quote was found on rather than a line number. A scanned page has no text layer, and says so instead of appearing blank. Saving opens a short form:- The selected text is shown but not editable — it’s the record, and it’s stored exactly as you selected it
- A caption is optional. Use it for what the text means; it becomes the headline for the claim, with the quote shown beneath it. The quote itself stays untouched, so an agent reading the claim can always tell your words from the source’s
- A recall priority controls how readily future agents see it, using the same scale as everything else in the graph. Each level explains itself in the picker; High additionally surfaces the claim to every agent in the workspace automatically
Provenance
Every claim links to its sources:- Notebook cells: Jump to the cell that produced evidence
- Files: Open the referenced file, scrolled to the quoted line range — or, for a PDF, opened to the page the quote was found on
- Other claims: Navigate the claim graph
How agents find past claims
Agents look up relevant findings from earlier work. High-priority claims are shared automatically with every agent in the workspace; other claims are retrieved as needed.Knowledge activity in chat
Chat activity shows when agents create, update, or search knowledge. Click a claim’s title to open it in Project, or expand a search row to inspect the results.Claims and tasks
The knowledge system connects to tasks:- Tasks require synthesized claims for completion
- Claims link back to the task that produced them
- The task graph and claim graph together document your research path
Task-variable integration
When tasks pass variables to downstream tasks via task-variable evidence (thequalia package), those values become runtime-captured claims. This creates a documented chain of data flow through your workspace.