HOW IT WORKS
The relay that remembers
One API endpoint. A full cognitive chain. Persistent memory that compounds with every invocation.
THE MENTAL MODEL
A Brain Instance is not a chatbot
Thalium is intelligence middleware. When your application sends input to a Brain Instance, it does not simply call a language model and return text. It runs a structured, multi-role cognitive chain — classifying intent, retrieving relevant memory, structuring a response, challenging it, gating on confidence, and writing the result back to the institutional ring.
The Brain is named after the thalamus: the brain's central relay nucleus. Every signal passes through it. Every signal is remembered.
THE INVOCATION LIFECYCLE
01
Invoke
Your application sends input, domain context, and an optional session ID. Thalium opens an Anchor — a Redis-backed chain ledger that is the single source of truth for the entire execution.
02
Triage & classify
A lightweight classifier assigns one of 11 intent types (specification, diagnosis, verification, risk assessment, and more) and derives an address key — the routing coordinate for memory retrieval.
03
Role chain executes
Up to 13 roles run in sequence. The Architect structures output. The Devil challenges it. The Scorer gates on confidence. The Validator checks for reclassification. Each role writes to the Anchor.
04
Consolidate & learn
The Librarian writes the artifact to the institutional ring. The Calibrator refines scoring rules from accumulated experience. The Coverage Map updates. The Brain is permanently smarter.
MEMORY ARCHITECTURE
Three rings. One address system.
Thalium's memory is structured, not vectorised. Every entry is filed at a deterministic address key derived from Triage classification. Retrieval is constant-time. There is no semantic drift.
Session ring
Live context for the current session. Resets on session close. Used for chain continuity within a single interaction.
Entity ring
History of a specific named thing — a system, a contract, a user, a supplier. Retrieved when that entity is referenced in input.
Institutional ring
Compounding domain knowledge. Permanent. Every invocation adds to it. The Calibrator refines the rules derived from it. This is what makes the Brain smarter over time.
THE ROLE CHAIN
Structured reasoning, not raw generation
Each role is a focused, independently testable function. No role knows more than it needs to. State travels in the Anchor.
RESPONSE TIMING
Two artifacts. One invocation.
Every invocation runs two parallel paths: a Fast Chain that returns within seconds, and a Full Chain that completes the deep reasoning. You receive both — your application can show immediate feedback, then update when the full result arrives.
1 — 3 seconds
p50: 1.5s · p95: 3s
Produced by the Fast Chain: Triage → Listener → Scorer only. Low-fidelity but real — carries the classification, address key, and a confidence score.
Use this to show immediate feedback to your user. Display a loading state, a preview, or an early signal — without waiting for the full chain.
6 — 15 seconds
p50: 6s · p95: 15s
Produced by the Full Chain: all active roles. Complete artifact with full provenance, anchor trace, model costs per role, and memory write confirmation.
Supersedes fast.artifact when it arrives. Update your UI with the definitive result. The institutional ring write is confirmed in this event.
SSE EVENT SEQUENCE
INVOCATION FLOW — LIVE
API CALL → ANCHOR → ROLE CHAIN → ARTIFACT → RING WRITE
THREE-RING MEMORY — LIVE
Memory that compounds
The innermost ring (session) resets after each session. The middle ring (entity) persists the history of named things. The outer ring (institutional) grows permanently — every invocation adds to it, and it never forgets.
Watch the outer ring brighten over time as entries accumulate. That brightening is the Brain getting smarter.