Cerevisor 4.0.0: Memory in layers, runs that check their work, GitHub Copilot

Memory now comes in layers: About you, My business and one page for each area of your work, and everything Cerevisor learns waits as a suggestion until you accept it. A run checks what its skills need before it starts, Cerevisor runs the checks you set on an output itself, and a run that leaves loose ends says so as Finished, needs attention. GitHub Copilot joins the providers. A workflow can now wait for a person, start from a webhook and tell another app what happened, and after a crash one window lists everything that was interrupted, each with its own Resume.

1. Memory in layers

Memory is now one line per fact, in three layers: About you, My business, and a page for each work context (an area of your work, such as a client or a project). They replace the single profile. Each agent gets the lines of the workflow’s own work context and none from the others. The Pre-Run Preview counts the lines that did not fit.

• Your profile moves over once. On first start, your existing profile and its open suggestions are copied into lines. The old profile file stays on disk, unchanged.
• Work contexts are created, ordered, archived and brought back on My business. File a workflow under one in Settings → Workflow → Options or on the Launcher. Sort my workflows suggests a work context for each unfiled workflow and files nothing until you press File N workflows. The AI Builder files a new workflow when one context clearly fits and says Filed under … · Change. If you choose No work context, that choice is remembered.
• Nothing is learned without your click. What Cerevisor learns from runs and chats arrives on Memory → Suggestions, with Accept, Edit and accept and Reject. A rejected line is not suggested again. When a suggestion contradicts one of your lines, the card shows This differs from what you said: …, and accepting it replaces your old line.
• Lines shared by two work contexts are offered to My business: Said in 2 work contexts: ….
• Lines that age. Right now and Under way lines fade after 14 days without a refresh. A goal nobody touched for 30 days comes back as Is ‘…’ still a goal? with Still a goal, Done and Pause.
• Learnings stay in their context. A learning saved in a run filed under a work context reaches only workflows in that context.
• Fill a page quickly. Each page has Answer a few questions (four short questions) and Start from a document: paste text, choose a file (Markdown, text, PDF, Word or PowerPoint), or type a web address and press Read. With two or more work contexts, Draft from my work contexts drafts My business. Every line is shown, editable, before anything is saved: Nothing is saved until you press Save.
• Delete on About you and My business is immediate, with Undo in the footer for a few seconds.

Where it goes. Your lines are kept on your computer. The ones that fit a workflow go to that workflow’s model provider, as part of the prompt. After a run or a chat, your default chat model reads the lines that fit it, to suggest what to remember. Reading a document or splitting your answers uses your default chat model too. Nothing is sent to Cerevisor. A web address is opened only because you asked.

Confirming a theme in Memory → Harness now takes about 0.2 seconds instead of 2.5, and the tab no longer goes blank while it reads again.

2. Runs check before they start and report what is left after

A skill says what it needs. A skill can list the programs, Vault secrets and connected tools it needs (mark one optional with a ?). Before a run, Cerevisor checks each agent that holds the skill. When something is missing, a card says This workflow may not be able to finish, with Fix and recheck and Run anyway. Secrets get a presence check only: no value is read and no service is called. To check a connected tool, Cerevisor starts its server as a run would, and waits at most 20 seconds.

Checks on an output. Under an output file’s contract, a Checks field takes up to 3 commands, one per line. Cerevisor runs them itself after the agent writes the file, and a failing check sends the agent back once. Checks run only when the workflow allows commands, and they get no Vault values. New or changed checks need your approval, like a new power. A shared workflow cannot carry checks: they are dropped on export, and a shared package that has them is refused.

Finished, needs attention. Every agent is asked to end its work with a short Not verified list. When a run leaves loose ends, it ends as Finished, needs attention, in amber, in the run window, the footer, the agent cards and the world. A close-out under the run’s results lists what the agents could not check, the checks that failed, and, in a git folder, the files the run changed and did not commit. Cerevisor reads git for this itself, read-only, even when your agents may not use git.

3. GitHub Copilot as a provider

An agent can now run on GitHub Copilot, through your own Copilot command-line program. Open the GitHub Copilot tile in the Provider Hub, marked Preview: it shows GitHub’s install commands, Sign in with GitHub, and an optional model (leave it empty for your Copilot default).

• Copilot asks Cerevisor before every risky step: a new file, a change, a command, a connected-tool call, a path outside the working folder, and the same for Copilot’s own helpers. Your approval mode answers, as for Cerevisor’s own agents.
• What the workflow switched off stays off through Copilot’s own rules, and a run stops if anything tries to switch Copilot’s permission checks off.
• For runs Cerevisor starts, Copilot’s telemetry and its automatic updates are off.

Copilot’s connection for outside apps is still in public preview at GitHub, so the tile says Preview. Each run is one Copilot session on your own plan and counts against your own limits.

4. Human steps: a run waits for you

A human step is a card where the run stops for a person. Add one from Manual controls → Add Human step, or ask the AI Builder.

• When the run reaches it, a panel called Your input opens with what the agents handed over. Answer on a plain form, or on a page the AI Builder designed for the workflow and saved in its folder. Pages open inside Cerevisor, never in your browser.
• The button you click picks the next step, and your answer is what the next agents read.
• It waits as long as it takes, in scheduled and Hands-free runs too. A waiting step calls no model and takes no place among the agents that may run at once. A desktop notification and the footer tell you a step is waiting; a paired phone gets a notification.
• A page cannot answer by itself: an answer needs your own click or key inside it. If anything other than the AI Builder changed the page, you see the plain form until you press Use this page.
• Several human steps in one wave show one after another. Settings → Admin → Show human steps side by side changes that.

5. Webhooks

Start a workflow from a URL. Choose Webhook on the Workflow Input card, or the From a webhook tab of the Schedule window. Cerevisor gives the workflow an address on this computer and a secret. What arrives is saved in an inbox folder inside the workflow folder and treated as outside content, like a web page. A delivery that arrives while the workflow runs waits its turn by default; you can choose to drop it instead.

The address is reachable from this computer only. To reach it from the internet, put your own tunnel in front of it.

Tell a URL what happened. Send to a URL is in Settings → Automation for every workflow, and in a workflow’s own settings as This workflow sends to. Each message is signed. A failed send is tried again after 1, 5 and 30 minutes. Send a test and a delivery log show what went out. A run that finishes with loose ends reports outcome: "completed_needs_attention" beside status: "completed".

A shared workflow carries no address, secret or subscription, only the fact that it needs them. Copy Workflow, Export and the workflow file do carry a workflow’s own send-to addresses, so look at them in a file someone else made.

6. Interrupted work comes back

After a crash, one Interrupted work window lists every workflow run and every chat reply that was cut off, newest first, each with Resume and Dismiss. A run row says what Resume keeps, continues and runs again, and View opens that run.

• Resume brings the workflow on screen and continues in one run: finished agents are kept, Claude Code, Codex and Cursor agents continue in their own session, and the rest run again from their start.
• Chat replies resume in the agent’s own session, and the agent is told to check what is already done before doing it again.
• The per-agent Resume agent is gone; it started a fresh run on top of the recovered one.

Ctrl+R no longer hides running agents. A window that reloads while runs go on catches up: agents at work, chat replies and waiting questions show again, and Stop still works. The window’s own errors are now written to Cerevisor’s log file on this computer, where the help chat can read them.

The crash screen says how to get going again: This window stopped working, then Ctrl+R (⌘R on a Mac) or Reload window. The error sits behind Technical details, with Copy details.

7. Notes on the canvas

Right-click empty space in a workflow and choose Add note. A− and A+ set the text size from 10 to 72 px, and a note comes in five soft colours: Plain, Yellow, Blue, Green and Pink. Resize, duplicate, copy and paste notes like cards. Notes sit behind the cards and are for people only: no run, prompt, AI Builder or phone reads them. A shared workflow carries its notes, and the Share dialog says so.

8. The Operator report

The Operator panel has two views, Report and Overview. Report opens first when a check finished since your last visit. Since you were last here gives the number of checks, what they spent and today’s spend against your cap, a sentence for each check that did something, and a Review line when something waits for you. Below it, every check is listed by day. Open one to see what it saw, its reasoning, what it did and what came out, with View run and Show on canvas. No model writes the report, and nothing leaves your computer.

9. Agents can call web APIs, and approving a site takes one click

The web fetch tool now handles an ordinary REST API:

• A request body, with Vault secrets filled in, as anywhere else.
• A time limit, 30 seconds by default, between 1 second and 5 minutes.
• Redirects checked at every hop. Before, only the first address was checked, so a site could redirect a call to this computer or off your approved list. A hop to another site now carries none of the agent’s keys, and a write is never sent on to another site.
• Large answers saved whole to a new file (with File creation on). A fetch never overwrites a file.
• A failed connection names its reason.

Approving a site. When you press Run or send from the launcher, Cerevisor reads the web addresses in the agents’ instructions. Sites not on your list show on one card, This workflow calls a new site, with Add and run, Run without adding and Cancel. When an agent asks for a site during a run, the prompt has Add … to approved sites. Your approved sites now count under Anywhere too, so an agent is no longer stopped on its second call to an API it was allowed to use.

10. A terminal for each workflow

A button right after the model chip in the Workflow bar opens a real shell in the workflow’s working folder: Command Prompt on Windows, your login shell on a Mac, your shell on Linux. The window moves and resizes inside Cerevisor, + opens more terminals as tabs, Hide keeps them running, and Close ends them. Every key goes to the shell, Ctrl+S included. The terminal is yours: it is not a run, no agent can see it or type in it, and nothing typed or printed is recorded.

11. Settings search

Settings is wider, with its 19 tabs in four groups down the left. Type in Search settings (Ctrl+F, or Cmd+F on a Mac) and the matching settings appear at once as the real controls, ready to change. Each says where it lives, whether it applies to This workflow or All workflows, and has an Open in … link. Esc clears the search, and a second Esc closes Settings. Field edges and faint text are easier to see across the app, in every colour scheme.

12. Connected tools: choose which tools, and T3 Code

• Choose which of a server’s tools agents may use, in the server’s Advanced settings. A tool you leave out never reaches any agent.
• Let tools that only read run without asking is a new switch per server. Their results are still treated as outside content.
• T3 Code has a card among the suggested connections. It asks for your own T3 Code address and starts with eight of its tools.

13. Smart routing that pays its way

• A cheaper tier must really cost less. Routing moves an agent to a lower tier only when that model costs less on every kind of token, and never picks a model that another model of the same or a higher tier on your account beats on every price.
• Route each agent in the workflow’s model menu picks a model for every agent at once, and one Undo puts them all back.
• What a run is worth. On a subscription, the run’s Overview shows its list price, included in your subscription, not charged. When the cache saved money, it shows Cache saved $….
• Analytics has a switch to show subscription runs at $0 or at list price, and so does the phone, in its Costs settings.
• On Claude Code, Cerevisor suggests turning routing on when it was off and an Auto agent ran on a powerful model.

14. The team conversation

• It keeps its history after a restart, for runs and chats: the newest 60 notes, with Show earlier notes. Summary gives a timeline of who decided what, and a written summary on request.
• Start fresh draws a line in the conversation. Every note stays to read, and no note from before the line reaches an agent again. Undo puts the line back.
• Writing to an agent by name wakes it. In agent chats, a note addressed to an agent starts its next turn, up to 8 times per workflow between two of your messages.
• Less repetition. Agents that continue a chat are told only the notes since their last turn.

15. Smaller things

• Copy and paste agent cards where you point: Ctrl+V and right-click Paste land under the pointer, in waves and on a free canvas. A card added by hand on a free canvas lands where you put it.
• Done and failed dots. Finished agents show as dots after the agents at work until you press Clear. Failed agents get red dots of their own; press one to go to the agent and see the error, and clear each one when you have dealt with it. A chat reply that ends in an error counts as failed.
• AI Builder dots. A line under the agent dots shows each AI Builder at work. Click one to open that builder.
• The launcher keeps what you type until you send it, through Escape, a workflow switch and a restart. A request sent while the workflow runs joins a queue and starts when the run finishes.
• The AI Builder works in the background, on a workflow that is not on screen. It fixes changes the canvas refused by itself, and keeps cards in their waves.
• A missing input says why. When a run cannot start because an input file is missing, the workflow frame shows the reason with a See why link to the Log, and names the agent that writes a file of that name and where. The AI Builder now keeps an output where it was asked, so its readers find it.
• Agent instructions stay a sensible length. The AI Builder keeps an agent’s instructions under a limit and offers to move long procedures into a skill. Set it in Settings → Chat Builder → Agent prompt length (800 words by default, 200 to 3,000, or no limit).
• Long builder chats cost less and stay fast. An older chat is shortened into a recap, or a Claude Code or Codex builder starts a fresh session, under Settings → Chat Builder → Long chats. The chat shows its latest 80 messages, with Show earlier messages.
• Stop & start this in the AI Builder works with one press.
• A decision card’s input and output chips open what they name, not the card’s settings.
• The help chat can look into your runs: why one failed, was slow or cost a lot. A Log details switch, off by default, lets it open single log entries in full.
• Point on screen works when the help chat runs on a subscription app.
• Zooming no longer flickers.
• No lines between workflows on the world canvas. A link between two workflows still works and shows as the shared file.
• The “Drop files to attach” overlay no longer stays stuck after a drag passes over the app.
• The Ocean look gives the guide’s buttons a readable label.

Changes to how every run behaves

• A run no longer opens its log by itself. Open it with Log in the footer. Permission prompts still show as before.
• A run with loose ends ends as Finished, needs attention. Every agent is asked for a Not verified list. A run in a git folder that changes files besides its declared outputs, and leaves them uncommitted, ends in amber. Expect more amber runs than before.
• Cerevisor reads git at the start and the end of a run, read-only, to build the close-out. A slow repository can add up to 10 seconds at each end.
• Skill requirements are checked before a run, and a connected tool a skill needs is started before Run.
• Memory reaches prompts as About you, My business and This work, and no learned line is added without your click.
• Your approved sites count under Anywhere too, and every redirect of a web fetch is checked again.

Security

• Electron 44.10.0 (Chromium 152.0.7977.169, Node 24.22.1), up from Electron 44.4.5 in 3.9.0. It brings Node’s October security release.
• A web fetch no longer hands your keys to another site. Before, a redirect from an approved site to another one carried the agent’s keys and request with it, and could reach this computer. Every hop is now checked, and a hop to another site carries none of them.
• A connected tool’s server that redirects to another site no longer passes that site your custom headers or request body (GHSA-6prh-2h8m-c8cw).
• No new kind of data leaves your computer. Memory goes only to the model provider of the workflow that uses it, and to your default chat model when it suggests what to remember. Webhooks listen on this computer only, and send only to addresses you set.
• Fixed:
• Electron: GHSA-x562-297c-g43g (a permission granted to the wrong frame).
• The connected-tools library: GHSA-6prh-2h8m-c8cw, and GHSA-46jr-4wx8-v453 for Cerevisor’s local API, which now refuses an oversized tool call.
• The code editor’s HTML cleaner, now really on the patched version: GHSA-cg2g-xh64-v4q7, GHSA-mrv2-3pqr-j5jr.
• Outlook sign-in library: GHSA-xcx6-jx8p-x3h5.
• Input checking library: GHSA-mrhx-c7xc-m3w5.
• A web request parser: GHSA-27r3-jxmh-p828, GHSA-v422-hmwv-36x6.
• Removed with the Cursor integration’s updated library: GHSA-x8mw-p69m-v3mx, GHSA-gxm5-99cw-xjw9, GHSA-vrm6-8vpv-qv8q, GHSA-v9p9-hfj2-hcw8, GHSA-vxpw-j846-p89q.
• Open: no high or critical report. Seven moderate and low reports remain in libraries Cerevisor carries (hono and @hono/node-server, fast-uri, ip-address, postcss, uuid, @babel/core). Whether Cerevisor reaches the affected parts has not been checked yet. No exception is on file: the list of accepted findings is empty.
• Windows installers are not yet code-signed, and macOS builds are not yet notarized.

Known limits

• Memory: how well a real model splits your answers, reads a document or drafts My business has not been tested yet. About you and My business reach every workflow, household work included, and learnings saved before 4.0 reach every workflow. After the update, an older version of Cerevisor can only read your memory, not change it.
• The close-out: a file already changed before the run and changed again by it is not listed. In a repository with Git LFS set up for that repository alone, the close-out says it did not read the git state. A paired phone shows Needs attention only after it refreshes.
• Checks on an output run without a prompt that names them on a workflow’s first run and in a workflow file you open from disk. Read the checks of any workflow you did not write.
• GitHub Copilot was tested with a stand-in model, not yet signed in to GitHub’s own models, and not on a Mac or Linux.
• Human steps: a page made by the AI Builder on a real model has not been tried. A phone gets the notification only with the updated companion app, and only while the app is open.
• Webhooks work only while Cerevisor is open with its window. They were tested on Windows, not with a real tunnel or a real Slack, Discord, n8n or Zapier receiver. Retries waiting when you quit are not sent.
• T3 Code: signing in to T3 Code from Cerevisor has not been tried end to end.
• Web APIs: no real model has used the new fetch options yet. A very large answer is read whole into memory before it is saved.
• Resume: an agent on Anthropic, an OpenAI-compatible provider or Gemini that was cut off mid-way runs again from its own start.
• The terminal has not been tried on a Mac or Linux, and links in its output cannot be clicked.
• Outlook and Cursor run on updated libraries (the Outlook sign-in library moved up two major versions). Both were checked by tests; signing in to Outlook and running a Cursor agent were not tried live for this release.
• Nothing in this release was tried on a Mac. It was built and checked on Windows.
• Everything listed under 3.1.0’s through 3.9.0’s known limits still applies.

Platforms

Available for Windows, macOS (Intel and Apple silicon) and Linux (AppImage and Debian package). Using apps is on Windows and on a Mac only; on Linux the setting does not appear.

Updating

On Windows and Linux, existing installations receive the update through Cerevisor’s built-in updater. On a Mac, download the new version from cerevisor.com/download and replace the app in Applications. Your workflows, run history and settings carry forward untouched.