v1.17.0 — Stop re-explaining your project
Every conversation used to start from nothing. You'd re-state the same things — who the client is, which style guide applies, what you already decided against — thread after thread, agent after agent.
Projects now have a memory.
Project memory
Open a project and you'll find a new Memory tab. Notes kept there are handed to every agent answering inside that project. No @mention, no re-pasting, no "as I mentioned earlier".
Each note is one of four kinds, because they're used differently:
| Kind |
For |
Example |
| Fact |
Something simply true |
"The client is a B2B logistics firm in Rotterdam." |
| Preference |
How you like things done |
"Always use metric units and ISO dates." |
| Decision |
Something already settled |
"We chose Postgres over DynamoDB in March." |
| Constraint |
A rule that must hold |
"Never mention the parent company by name." |
Notes run to 8,000 characters and arrive verbatim — what you wrote is what the agent reads, not a summary of it.
The whole loop — a note can start with you or with a finished conversation, and you decide how much the automatic half is allowed to do:
graph TD
W["You write a note"] --> T["Memory tab"]
C["A conversation finishes"] --> X["Background pass reads it"]
X --> M{"Capture mode"}
M -->|Automatic| T
M -->|Suggest| S["Waits for you to accept"]
M -->|Off| N["Nothing captured"]
S --> T
T --> A["Every agent in this project"]
A --> D["Reply names what it used"]
It fills itself in
You don't have to write them. As conversations in a project finish, a background pass reads them and proposes notes worth keeping. You set how much rope that gets, per project: Automatic (added as found), Suggest (proposed, waits for you), or Off (the tab is yours alone).
Anything captured automatically is labelled as such and links back to the conversation it came from.
You decide what always makes the cut
A long project can hold more notes than one turn has room for. So memory is layered — Always (pinned, in every time), Relevant (pulled in when they match the question), and Automatic.
Pinned notes are dragged into the order you want, and that is exactly the order they're seated in. When a turn genuinely can't hold everything, what you ranked last is what gets cut.
You can see what the agent was told
Every reply that used project memory says so and names what it was given. If an answer surprises you, check what it was working from instead of guessing.
Memory belongs to the project, not to you — any member can read and edit it.
Workflows tell you the rules while you build
The builder used to let you assemble something that could never run and only say so at run time.
- A workflow that cannot run will not publish. Wire a step to a tool its model can never be handed and publishing is refused, with the problem named.
- Every published workflow has its own link, and if it needs inputs first it shows you its own form instead of an error.
- Checks that lived only in the executor now run in the editor.
Three ways to feed a workflow
Chat-invoked workflows take their source from whichever is nearest: paste it, drop a link, or use the conversation you're in.
That last one is deliberately bounded — a workflow reads the thread only when its author opted in and only for a workflow you own. Being able to run something isn't permission to read your conversation.
New template: Social Kit — one source in, per-network copy plus correctly sized visuals out, for eight networks.
Routing keeps getting sharper
- "Hey there" no longer picks a roleplay persona. A greeting names no task, so searching the roster for one returned confident nonsense — the general-purpose agent ranked last of eight. Small talk goes to the generalist now.
- A tool an agent was born with no longer counts as a choice. Some tools come automatically with the model an agent runs on — everyone in that family has them — yet they were scored as distinguishing skills. The worst case: the image-description tool ranked highest of all auto-granted tools, and it's given to exactly those agents whose model cannot see images. Image questions were being nudged toward agents blind to them. Only deliberately bound tools count now.
Security
An account-takeover path is closed: a flaw in how anonymous sessions were identified could let someone reach another person's account without knowing anything about it. Anonymous identity was rebuilt and the old mechanism removed entirely.
The full release note lives in the in-app What's New drawer.
Tell us what to build next in Feature Requests.