Part of The Always Verify Stack — a matched three-layer set (AIDL · CogStack · EthicalSwarm) for governing multi-agent, multi-vendor AI.
A two-layer architecture for autonomous AI agent systems: a persistent cognition substrate that gives agents memory across sessions, and a multi-vendor cognitive stacking model that assembles agents from different providers to produce emergent reliability through decorrelated diversity.
Current AI agents are stateless — context evaporates at session end. And naive multi-agent setups run the same model multiple times and call it redundancy, when what they've actually built is correlated agreement dressed up as consensus. CogStack addresses both: externalize and rehydrate the context so agents reason continuously, and stack agents from genuinely different providers so the composite covers failure modes no single constituent has.
| File | What it is |
|---|---|
SPEC.md |
Technical architecture: persistent cognition substrate, cognitive stacking topologies, AIDL integration |
DISCLOSURE.md |
Defensive publication document |
PROVENANCE.md |
Origin, timeline, and why this is published openly |
Layer 1 — Persistent cognition substrate. A Context Compression Engine compacts interactive session artifacts into typed semantic blocks. A Memory Store holds them durably. A Rehydration Pipeline injects relevant blocks into subsequent API prompts within the model's context budget. An Execution Daemon runs headlessly, polling a Directive Manager for tasks and appending every action to an Audit Log.
Layer 2 — Multi-vendor cognitive stacking. Agents are drawn from different providers and model families specifically to maximize genuine independence — not nominal vendor diversity. Four composition topologies are defined: parallel-and-adjudicate, sequential/pipeline, adversarial, and router/dispatch. Stacking is invoked selectively for high-stakes, correctness-critical, or anti-sycophancy-requiring tasks; it is not the default path.
CogStack is the orchestration layer. AIDL provides the governance grammar that bounds it: agent roles, reputation accumulation, and adjudication weighted by provider independence. EthicalSwarm provides the ethical governance framework for multi-agent deployments at scale.
The three form a matched set: CogStack orchestrates, AIDL governs, EthicalSwarm bounds the ethical perimeter.
Published openly as a contribution and as prior art, under the MIT license. This is a specification and reference design — not a product. A consolidated technical disclosure is archived as CogStack-Defensive-Publication.pdf and submitted as a defensive publication on the Technical Disclosure Commons.
MIT. See LICENSE.