Remember what the project knows
Your agent should not start from zero every time.
Paste everything into chat and the model drifts. Workbench keeps a living memory of your project so the agent stays oriented without dumping all your notes into every turn.
Today, most AI coding sessions forget the architecture the moment the chat ends.
Session
Where the work happens. You steer. The agent ships the change.
Knowledge
Your notes. Open, fix, and keep them forever yours.
Brain
The map. See how the pieces connect before you dig in.
The notes
Knowledge
Your notes the agent cannot rewrite without you.
Open your project notes. Fix a wrong fact, add what only your team knows, and keep company, org, and repo memory in one place you control.
- Correct a wrong fact the moment you see it
- Keep shared memory for the team, not buried in one chat
- Sync notes when you link a remote, on your terms
Memory lives at every layer
Company
Defaults and playbooks everyone starts from
Team
Shared rules across the repos you own
This repo
What is true for this codebase alone
Knowledge fields
How Knowledge fields shape the map
Workbench reads a few fields at the top of each Knowledge page so it knows what the entry represents, how strongly it should matter, and how it connects on the map.
| Field | What it means |
|---|---|
| Type | What kind of thing this is a concept, a service, a repo, a rule, and so on. |
| Id / Title | A stable id so the map does not break when you rename, plus the name people see. |
| Role | Hub (a main concern), spoke (something that hangs off a hub), or a plain entry. |
| Aliases | Other words people type for the same thing so “frontend” and “ui” can land on one hub. |
| Tags | Soft hints that nudge retrieval when the wording is close but not exact. |
| Salience | How important this entry is (0 to 1) when ranking what to show the agent. |
| Primacy | A boost for foundational hubs the stuff the project is built on. |
| Sources | Paths that were inspected when this structural entry was created. |
| Edges | Typed links to other entries related to, part of, uses, contradicts, and similar. |
| Conflicts with | Explicit “this disagrees with that” targets so contradictions stay visible. |
Sketch
Start from the repos in your system
Deepen
Fill in the important hubs and connections
Explore
Ask about a concern and follow the trail
The map
Brain
See the project. Then ask around a concern.
Brain turns your system into a readable map services, standards, frontend, devops, and the links between them. Grow it once, then explore a question without losing the big picture.
- See how frontend, backend, and ops actually connect
- Grow the map before a big feature, not after the mess
- Ask around a topic and follow the path that matters
Two surfaces. One memory.
Knowledge is where you read and write. If the map is wrong, you fix the entry.
Brain is where you look. If you need orientation, you open the map.
Two ways in
Build memory before the rush
Teach the project
- 1.Open Brain and pick the system you care about
- 2.Grow the map until the big pieces show up
- 3.Explore a topic frontend, deploy, standards
- 4.Fix anything wrong in Knowledge
Stay in the session
Just ship
- 1.Start Work like you always do
- 2.The agent gets only the notes that fit this job
- 3.Save important facts back when they matter
- 4.Open Knowledge later if something feels off
While you work, the right notes show up.
Sessions pull a small, focused slice of memory not the whole set of notes. When something important lands, save it back so the next session starts smarter.