The Stream is what's new. The Atlas is what's connected. Knowledge Atlas links decisions, lessons, code, docs, tasks, sessions, and project history into a queryable graph so agents find the right context, not just the most recent context.
Most "memory" tools store snippets. The Atlas stores relationships — so when an agent reads auth/session.ts, it also sees the decision that shaped it, the lesson that constrains it, and the open task that depends on it.
Every file gains the decisions, trade-offs, and lessons that shaped it. No more re-deriving why.
Lessons promote into guardrails on the surfaces they protect. Same fact, two roles.
Indexed READMEs and runbooks attach to the sessions that actually used them.
The relationships between files, concepts, decisions, and lessons live in one graph. Toggle the view; the underlying truth is the same.
Imports, references, and ownership graphs derived from the source — sanitised, never stored.
Wisdom nodes connected by who learned them, where they were used, and which code they protect.
MCP-compatible tools query the Atlas with scoped, source-backed responses.
The first-party CLI uses the Atlas before every edit, surfacing related context inline.
Capsules are pre-traversed Atlas slices, scoped, expiring, and revocable.
Two views of the same graph: the knowledge graph (decisions, lessons, docs, sessions) and the code graph (files, modules, symbols). Both queryable from any MCP-compatible AI tool.
The Stream is what's new. The Atlas is what's connected. Start with one decision; the graph grows from there.