Agent Brains and Control Loops
Why 'agent brain' means two different products — memory wikis versus air-traffic control — and why the control-loop half is where fleets live or die.
Last updated: Aug 22, 2026The market collapsed two different products under one word. A memory brain persists world knowledge; an ATC brain ranks work, assigns idle harnesses, and rests when nothing warrants action. Buying or building the first does not give you the second, and wiring a wiki in as your controller recreates the exact failure loops we documented at 1.191 billion tokens.
The split that matters
Two products share the word “brain” and must not be collapsed:
| Class | Job | Examples |
|---|---|---|
| Memory / wiki-brain | Persist, retrieve, synthesize world knowledge across sessions | GBrain, Hindsight, Mem0, Zep/Graphiti, Letta archival |
| ATC / ops-brain | Rank work, pick idle harness, gate actuation, rest when unchanged | Mumega brain + cortex + work-router |
Industry consensus in 2026 is blunt: memory ≠ orchestration. Frameworks don’t solve cross-session institutional memory; memory products don’t solve idle-harness assignment. The failure mode to avoid is “ask the wiki what to do” — an always-on enrichment loop as the primary controller has both maximal cost and maximal hallucination surface, positioned exactly where fabrication becomes fleet churn.
What the ATC half must guarantee
Our production incidents reduced the requirements to three properties (developed fully in Idempotent Agent Control):
- Decision idempotence — re-applying a supervisor decision to an already-satisfied state is a no-op, not another billion tokens.
- Cite-or-refuse — every actuation carries evidence resolvable by a cheap probe; failed resolution produces a logged refusal. Rest is valid — refusal as first-class outcome removes the incentive to invent work.
- Marginal-novelty gating — a new lane must buy a claim the last k lanes didn’t; otherwise the mandatory output is a coverage matrix, closeout, gap list, or redirect.
Underneath all three sits a restricted action space: {assign, skip, escalate}. Restart/deploy/merge/rotate are reachable only through escalation to a gated principal.
Where memory brains genuinely fit
We run read-mostly memory packs under our operator seats for world-facts (people, tenants, deals) via MCP, with writes restricted to assigned pages carrying citations. Overnight consolidation — Letta’s sleep-time compute, GBrain’s cron enrichment — is worth adapting as consolidation and re-ranking, never as a second controller with actuator access.
The one-line rule: memory advises; control decides; gates dispose.
Field notes
- What 1.191B Tokens Taught Us About Agent Supervision — the continuation loop postmortem
- What GBrain Teaches Us About Agent Memory
- Idempotent Agent Control (research) — the formal discipline
- The Failure-Mode Phase Transition — the degradation cousin of this failure family