Three layers of knowledge, one curated package per demand
A project's knowledge is rules, an index of its code and a memory of past demands. Each demand receives a curated package assembled from them, and the whole shelf is cloned into the sandbox for the agent to open what it needs.
What was on the table
An agent’s usefulness depends on the project’s rules, the map of its code and the memory of what previous demands found. That knowledge has to be permissioned per account and project, reachable from the sandbox, and — the part that decides quality — assembled per demand rather than dumped.
The paths we weighed
A raw document folder as the only mechanism. Rejected: with no curation the agent digs.
Everything in the prompt. Rejected: it grows with the project, not the demand.
An external RAG service per customer. Deferred; the port allows it later.
What we chose, and why
A knowledge base per project, in three layers: rules — the
conventions the agent obeys; index — one map per repository: what lives
where, how to build, how to test; memory — findings and lessons from
past demands. Text lives in the project’s root repository, under rules/,
index/, memory/ and demand/<id>/; binary artifacts live in the
object store, referenced from the repository.
A context package per demand is assembled when the sandbox is provisioned — the spec, the rules, the index of the repositories involved, the relevant memories — serialised deterministically so the prompt’s prefix stays cacheable. The package is curated; the repository is the shelf. The whole root repository is cloned into the sandbox with a generated manifest, so the agent opens what it needs without paying for it on every turn.
Write-back at closing: a demand’s findings and lessons are committed
into memory/. The index is regenerated on a merge event, not on a
schedule.
What it cost
Assembling the package and judging a memory’s relevance is the orchestrator’s work. Per-account, per-project permissions on the storage are covered by the repository and object-store contract suites.
Since then
Text moved from the object store to the project’s root repository on 2026-09-03, when the knowledge repository was decided; the three layers and the package were unchanged.