Workspace Memory
Persistent context that lives next to your code, merges across team and personal layers, and stays under human review. So your agents stop forgetting what they learned yesterday.
Memory that behaves like part of the codebase.
Travels with the code
Memory lives inside the repo it describes. Move the folder, branch it, share it — the context comes with it.
Layered context
Org, project, user, and local memory merge into every prompt. Personal notes never leak into the team's shared knowledge.
Human-in-the-loop
Agents propose; you approve. Nothing committed gets silently rewritten. Routine session logs auto-save in the background.
Built for swarms
Worktree-safe and orchestrator-mediated, so concurrent agents can read shared context without trampling each other's work.
One context, four sources.
Each tier covers a different scope. They merge in order — broader context first, more specific context last — so personal preferences can refine team standards without overwriting them.
Agents draft. You decide.
Auto-saving everything an agent says is how memory gets noisy fast. We split it: chat sessions log themselves, but anything that touches committed knowledge goes through a review queue.
Agent proposes
When an agent learns something worth keeping — a decision, a pattern, a gotcha — it drafts a memory entry instead of editing files directly.
You review
Proposals land in a queue with full context. Accept as-is, edit and accept, or dismiss — your call.
Memory commits
Approved entries land in the right tier, tagged and indexed. Next agent that opens the workspace inherits the new context automatically.
- Memory moves with the folder — no central registry.
- Loaded straight into context — no vector database to manage.
- Personal layer is gitignored — your notes stay yours.
- Concurrent agents are isolated by worktree, never racing on memory.
Coming to DevMesh soon
Workspace Memory is being built into DevMesh now. Get the desktop app to use it the moment it ships.