Architecture
How it all connects
The interesting part isn't any single app — it's that the agents, the websites, the mobile app, and the human all share one nervous system. Content, agent work, and human judgment move through the same loop.
Ecosystem map
The hub
Open Brain: one governed memory
Open Brain is a Postgres database with vector search (Supabase + pgvector) fronted only by edge functions. Durable content lives in a thoughts table; agent workflow memory lives in a separate eight-table governed sidecar with workspaces, visibility scopes, recall traces, and an audit log. Every agent, every automation script, and every coding session reads and writes through the same two endpoints: /recall before work starts, /writeback when it ends.
Governed writes
The writeback endpoint rejects credential-shaped strings, large code dumps, and raw conversation transcripts before they can become memory. A database constraint prevents machine-generated notes from ever being treated as instructions.
Recall observability
Every recall writes a trace: what was retrieved, its rank and similarity, and — after the work — which memories were actually used. Memory quality is measured, not assumed.
Hybrid search
Full-text search with web-style operators falls back to trigram matching only when needed, and ranking blends text relevance with importance and quality scores.
Per-agent silos
Each agent writes to its own memory partition, so recall never pulls another agent's noise. A cross-pollination bridge promotes only high-signal findings between silos — no direct agent-to-agent calls.
Coordination
Routing lives in the database
There is no orchestrator process. Agent pipelines are rows in Postgres: a task queue with an atomic claim-lock, per-agent workers that wake every minute, and a single database trigger that promotes the next stage when a task completes. Safety rails — a hop-count cap, a circuit breaker on repeated reclaims, and a rule that any unmatched route blocks for a human instead of guessing — are part of the schema, not the model's good behavior.
The design principle: specs are assets; runtimes are plumbing. Agent identities, memories, and pipelines live in files and database rows that would survive replacing the entire AI stack. Models are referenced by cost tier, never by name, so a provider swap is a one-file edit.
The memory loop · repeats continuously
01 · Content
Amy publishes a draft in Sanity or the Cue CMS
02 · /writeback
Governance filters reject credentials, code dumps, raw transcripts
03 · Memory
One superseding note per document, embedded for search
04 · /recall
Agents pull scoped, re-ranked context before they start work
05 · Agents
Drafts and reviews that have already read the latest content
06 · Amy
Her feedback is written back — the loop closes
Say hello
Curious about any of this?
No pitch — the work is the pitch. If something here sparked an idea, or you'd like a closer look at how a piece of it fits together, I'd genuinely love to talk about it.
Say hello