Loading
Loading
Release · August 19, 2026 · All news
This release opens the runtime to your own model providers and closes two long-standing gaps in governed memory: skill drafts that quietly die inside one session, and value claims that nothing actually proves.
Already installed? Ask your agent to “update Cognition” in plain English.
01
Cognition now supports customer-configured model providers with deterministic fallback paths, so the platform aligns with your model choice, cost posture, data governance, and existing enterprise agreements.
A complete deployment defines which provider receives each data class, what source content may enter a prompt, how prompts are minimized, whether providers retain or train on submitted data, how timeouts and failures are handled, what administrators can configure, and what is logged or redacted. BYO-model is a governed routing capability, not a provider dropdown.
02
Useful experience should not become durable organizational memory without review — but it should not disappear before review can occur. The checkpoint surfaces unresolved skill and workflow drafts at the end of a task instead of letting them die inside one session.
When should_render is true, your agent shows one approval card: what is pending your decision, what contradicts stored memory, what was applied, what was deliberately not used, and what value is actually proven. Nothing is approved on your behalf.
03
Saved-token claims now stay at zero unless an impact receipt, skill outcome, session receipt, or eval supplies proof. Transport and context trimming is reported separately, so value numbers can no longer be inflated by payload optimization.
estimated_tokens_saved = proven value only · estimated_payload_tokens_avoided = transport savings, reported apart.
04
Every saved skill now writes a supply-chain receipt tying its exact content hash to privacy, policy, eval, trust-gate, and route-policy state. Skill evals return signed, metadata-only traces with causal with-and-without claims, and team promotion fails closed until the gates pass.
Bitemporal replay can reconstruct what Cognition knew as of any recorded time while preserving superseded history. Raw prompts, raw failure bodies, and raw skill bodies are never stored in these receipts.
The hosted behavior is live now. Ask your agent to update Cognition in plain English, then get back to work.
