News
What changed, and what it proves.
Release · October 14, 2026Cognition 1.7.6: checks that know their limitsClinical checks now state how often a missed problem can get past a lighter review, and send a note to full review when it is unlike the notes that bound was measured on or when monitored rates drift. New claim-level checks cover citation support, dose re-derivation and missing inputs. Every note is still signed by a clinician.Release · October 14, 2026Cognition 1.7.5: fewer false gapsClinical checks no longer flag a covered requirement as missing because it was written in ordinary words, and generic words still do not count as coverage. The change was confirmed on documents the checks were never tuned on.Release · October 13, 2026Cognition 1.7.4: clinical checks pick themselvesEvery clinical note is matched to the right one of thirty workflow checklists from its codes or its own structure, with nothing to configure, and a note that fits none is held for a person instead of being checked against a guess.Release · October 4, 2026Cognition 1.7.3: your brain stays yoursA key stays attached to the workspace holding your memory. Recovery refuses a move that would hide it, access fails closed when a role cannot be confirmed, and executable skills run only in an isolated sandbox.Release · October 2, 2026Cognition 1.7.2: each release describes itselfThe installed instruction block now describes the release it announces. It had paired each new title with an older release's notes, so six releases told agents they were about something else.Release · September 30, 2026Cognition 1.7.1: wrong here is not wrongA recalled memory that does not belong is reported as irrelevant here, not incorrect, so it stops surfacing on the wrong kind of task without retiring it everywhere. Rules block only on a direct, in-scope match.Release · September 28, 2026Cognition 1.7.0: one answer per keyA workspace had more than one current state because readers each picked a different save. There is now one definition of what a key currently is, and a save that would empty a field it was not sent refuses instead of wiping it.Release · September 26, 2026Cognition 1.6.99: refusals name the fieldA gate refusal that named no field was not actionable. Refusals now name both fields and what each is for. Practice rules that could never be enforced moved into the channel that can check them.Release · September 24, 2026Cognition 1.6.98: silent success is a failureThree ways the product could report success while doing nothing: a gate with no steps answering safe, a memory tool that was not remembering, and a memory write with no telemetry. All three fail closed or record both outcomes now.Release · September 22, 2026Cognition 1.6.97: Cognition is the system of recordAgents are told durable memory lives in Cognition, not a local file. Revoking a key is no longer a removal: removal denies the identity at mint, auth, and renewal. A standing always-rule is no longer labelled never.Release · September 20, 2026Cognition 1.6.96: preferences land the same way twiceA stated preference had no route, so the same sentence did not land the same way twice. Unconditional wording applies every session, a condition becomes a skill trigger, and a one-off is deliberately not captured.Release · September 18, 2026Cognition 1.6.95: clinical documents under the same rulesTwenty-five clinical workflow registries, an HTTP plan gate with versioned policy packs, reviewer sign-off bound to the exact version reviewed, and an in-boundary runner that never sends chart text off the customer's boundary.Release · September 16, 2026Cognition 1.6.94: measurement catches up to judgmentFour defects found by reading one day's receipts against what happened: host markup was filed as judgment, a tie-breaker fired on every call, a retrieval miss counted as a skill failure, and a demonstrated win credited zero.Release · September 13, 2026Cognition 1.6.93: a brain can no longer be lost on joinA join that found nothing was creating something. Failed lookups and writes fail closed. An expired key renews instead of minting. The installer checks identity before writing, and the re-install loop that stranded brains is closed.Release · September 11, 2026Cognition 1.6.92: score the definition, not the numberWhen an agent answers an analytics question, the risk is a confident number under a definition nobody agreed to. Cognition now scores which reading an answer depends on, declares it, and refuses to pick when experts disagree or nobody has attested one.Release · August 26, 2026Cognition 1.6.85: opt-in telemetry, an uninstaller, and five bugs that broke your configSession-content telemetry is opt-in, disclosure comes before the first write, every shared config is backed up and never overwritten when unreadable, and one command uninstalls Cognition from every host. Re-run your installer to get it.Release · August 24, 2026Cognition 1.6.84: a plan it can ground, a done-claim it can proveSupplied plans built only from stored decisions, completion gated on evidence instead of confident wording, gates that vouch only for rules they can enforce, retrieval that finds the skill you wrote, and capture that scores effort over frequency.Release · August 21, 2026Cognition 1.6.78: the checkpoint card stops being a transcriptCompact audit checkpoint cards: summary caps at three bullets, proof collapses into one Verified line, safety-gated omissions read as Blocked safely, skipped checks as Not checked, and the long-audit shape is regression-locked.Release · August 19, 2026Cognition 1.6.75: bring your own model, keep the proofOpen-model runtime adapters, the approval checkpoint that stops skill drafts from dying inside one session, and proof-bound value accounting.