Brain
The fact kinds that ship, how facts get in, how recall picks them, and where to look in the console.
On this page
TL;DR: The brain stores facts about projects and recalls them into agent
runs. It is on in the local profile only; the production profile has no
brain: block, so it is off there.
Fact kinds
Declared in services/brain/: decision, delivery_risk,
engagement_phase and team_assignment, one workflow profile (delivery),
topics, importance weights, retrieval settings and synonyms. These are a
placeholder vocabulary; engineering kinds are still to be specified.
Apply changes through the brain config plane:
systemprompt admin config planes diff --plane brain
systemprompt admin config apply --plane brain --reason "..."
Entering facts
There are no extractors. Facts enter through POST /api/v1/brain/facts,
which needs no model.
Recall
services/brain/retrieval.yaml weights the tiers (pinned, subject,
linked, context, full_text, extension) and caps a recall at
max_facts: 12. The extension tier is the dense tier
(services/brain/dense.yaml), whose default embedder: hashing is
deterministic and needs no embeddings endpoint.
Console
/admin/brain shows Overview (kind health, recall per workflow, recent
recalls, disputes, consumer failures), Evals (stored recall eval runs) and
Sync (the declarations against the rows).
Recall quality on real model output is unproven.