# Workspace memory & procedures Two things carry over from one agent conversation to the next: **notes about your paper** that every new conversation reads, and **procedures**, reusable methods you can ask the agent to follow. Notes live in **Workspace memory**, beside your saved layouts. Procedures live in **Settings → AI & agent → Procedures**. Both are saved with the paper, so they survive restarts and travel with the project folder. The agent panel itself is just the conversation. Anything the agent saves along the way shows up as a card in the transcript. ## Notes the agent reads Open **Workspace memory** from the **Workspace** gear menu, or search for it with **⌘K**. The section **What the agent knows about this paper** lists notes about terminology, notation, datasets, decisions, constraints and open questions. A note can apply to the whole paper or to one source file. A file-scoped note is offered to the agent only when that file is in play. Each note keeps its source or reason: a lab-meeting decision, a passage in a project file or a page of a paper in your reference library. Open a note to read its full text, its source and when it was proposed. Keep writing preferences and project facts apart. “Prefer concise captions” is a preference and belongs in **Settings → Writing preferences**. “Experiment B uses split X” is a fact about the paper, so it belongs here, with the evidence for it. ## Approve, edit and forget notes The agent can propose a note while it works, but a proposal is not used until you approve it. Proposals sit at the top of the list marked **Awaiting your approval**. You can also approve one straight from its card in the conversation where it was proposed. 1. Open the proposal and read the wording and its source. 2. **Approve note** if it is right, or **Edit** it first. Saving your own wording approves it. 3. **Forget** a note you no longer want. Forgetting is reversible: **Forgotten notes** lists them, each with **Restore**. **＋ Add a note** writes one yourself. Your own notes are approved as soon as you save them. When the source behind a note changes, the note is marked **Needs recheck** and left out of new conversations until it is reassessed. Ask the agent to recheck it against the current source, or forget it. An early misunderstanding that becomes a durable note would be repeated in every later conversation, so read proposals before approving them. ## Build reusable procedures A procedure describes a repeatable method: the inputs it needs, its steps, the outputs to expect, how to check the result and any resources to consult. Open **Settings → AI & agent → Procedures** in a paper. Start from a built-in procedure (a rebuttal, a camera-ready review, a statistics audit or a related-work synthesis), write your own with **New procedure** or **Import SKILL.md** to review a Markdown draft. 1. Name the task and list the inputs it should ask for. Mark the ones it cannot do without as required. 2. Write steps that say what evidence to inspect and how to handle missing information. 3. Define concrete outputs and completion checks. 4. Choose the tool scope. **Research only** procedures cannot edit manuscript files. **Guarded manuscript edits** keep the ordinary file scope and approval rules. 5. Save. A procedure the agent proposes must be approved here before it can run. Resources are references for the agent to read. They are never run as scripts. ## Run a procedure in a conversation In the agent composer, open **＋ → Use a procedure** and pick an approved procedure. If it asks for inputs, a short form opens above the composer; fill it in and choose **Use procedure**. A chip above the message box shows the procedure and its version. **Inputs** reopens the form, and **×** removes the procedure. The procedure runs with your next message, pinned to the version you chose. Editing the procedure later saves a new version and does not change a run that already used the old one. If the procedure changed after you picked it, sending is refused until you pick it again. ## Versions, export and import Every save keeps the previous version. **Version history** lists them; **Use as a new version** starts an edit from an older one. **Export SKILL.md** downloads the current version as Markdown that other tools can read, and **Import SKILL.md** brings one back as a draft to review. **Archive** hides a procedure without deleting it; **Archived** lists archived procedures with **Restore**. ## What the agent saves during a conversation Besides notes and procedures, the agent can save evidence assessments for claims, literature comparison tables, revision plans and completion checklists. Each appears as a card under the reply that saved it. A saved claim's card opens its passage in the manuscript. The agent can read these records back in later conversations: ask it to summarize or update one. To check a claim against its sources, see [evidence](evidence.md). For numbers that must stay connected to their data, see [Linked results](linked-results.md). For a plan of the paper's sections, see the [manuscript board](manuscript-board.md). ## Where it is stored Notes, procedures and the other agent records live in `.versofold/research.json` inside the project folder. Layouts and reading positions in Workspace memory stay on this device. Writes are serialized and check each record's revision: if the agent changes a note while you edit it, saving is refused and the list reloads with the newer version, so you can reapply your change to it. These records are evidence and plans for the agent to read. They never grant it extra permissions. Related: [working with the agent](agent.md), [home & workspaces](workspaces.md), [find your way around](navigation.md).