Architecture
How Cognitive Runtime works.
A deterministic execution layer sits between your intent and the model. It defines what runs, in what order, against what rules — and keeps a complete record of every decision.
Execution model
Workflows as directed acyclic graphs
Instead of a prompt chain that the model navigates by feel, Cognitive Runtime represents every workflow as a DAG with explicit node dependencies. The execution engine supports node-level retry and idempotent execution, so a node can fail and re-run without corrupting state. The graph is the source of truth — not the model's interpretation of a paragraph of instructions.
Every run is recorded in an append-only event log — a never-modified sequence of what happened. That makes execution ordered, inspectable, and reproducible: you can trace exactly why a node failed and replay from that point.
Memory
Three layers of structured memory
Cognitive Runtime defines three memory layers with explicit scope and lifetime. Every read and write is typed — the system knows which layer a piece of information lives in and enforces access rules accordingly.
Episodic
Specific past interactions, stored in PostgreSQL. The record of what actually happened across prior runs.
Semantic
Learned patterns and embeddings, stored in a Qdrant vector database. Retrievable knowledge the system reasons over.
Procedural
Repeatable processes and routines, stored in PostgreSQL. The reusable know-how the system applies to recurring tasks.
Design principles
Built on six guarantees
Deterministic DAG execution
Every workflow is a directed acyclic graph with explicit node dependencies, node-level retry, and idempotent execution. Every run is backed by an append-only event log — never modified, fully auditable.
Layered memory (M1–M3)
Episodic memory in PostgreSQL, semantic memory in a Qdrant vector store, and procedural memory for repeatable routines. Each layer has defined scope and rules.
Meta-planning
Before a run, the Meta-Planner deterministically decides whether to REUSE a prior DAG, MODIFY an existing one, or SYNTHESIZE a new plan — rather than improvising step by step.
Reward & trust
A 7-signal composite reward score evaluates each decision, and EMA-based trust tracks agent reliability over time. Low-trust paths surface for review automatically.
Policy enforcement
Policy gates apply deterministic verdict evaluation at the runtime layer, not the prompt layer. Guardrails that can't be talked around, plus run cancellation, timeouts, and stale-heartbeat detection.
Code-change workflows
Structured DAG nodes for code modification, review, and decision control — reproducible engineering processes rather than one-shot prompts.
Why it matters
The reliability problem with prompt-based AI
When AI logic lives in a prompt, every output is probabilistic. A model that works 90% of the time sounds acceptable — until you realise that five sequential steps at 90% accuracy each gives you a 59% success rate end-to-end. Production AI systems need engineering guarantees, not statistical ones.
Cognitive Runtime separates what the model decides from how the system executes. The model handles reasoning — the runtime handles orchestration, validation, memory, and policy. That boundary is what makes reliable AI systems possible.