Skip to main content

October 2026

Covers: releases from 1 to 12 October 2026

Workspace admins can now have chats, reactions and reports expire automatically, and personas that stand for a real person can be grounded in what that person has actually said. Personas now follow their tuning notes and read recent writing on the topic at hand, and the quick scorecard on the home page picks a whole saved panel in one click. Kumkuat now publishes a privacy policy, terms of service and a Trust & Security section. For developers, the API gained canonical resource paths and a more readable reference, and agents that send a persona's name instead of its id get an answer rather than an error.

Data retention for chats, reactions and reports​

  • Set a retention window. A workspace admin can turn on automatic deletion under Workspace settings → Data retention: choose a window between 7 and 3650 days and which content it covers — chats, audience reactions and scorecards, narrative test reports, or any combination.
  • A background sweep deletes content older than the window about once an hour. Uploaded documents are not affected.
  • Off by default. Nothing changes for a workspace until an admin turns it on, and turning it off keeps content for the life of the subscription again.

See Data retention.

What a named person has said, on their persona​

For personas that stand for one real person — an analyst, an executive, a reporter — Kumkuat can now find where that person has spoken or written in your workspace's sources and attach it to the persona.

  • Find what they've said. In the persona editor, under Writing samples, Find what they've said lists the statements attributed to the person, newest first, with the documents they come from. Tick the ones to keep and choose Attach selected & tune voice (or Attach only).
  • Held to the record. Attached statements are given to the persona as verified public statements, so chats and Content Writing stay close to what the person actually said, and the persona's profile shows What they've said.
  • Where they stand. The same personas get a Grounding questions section. Kumkuat can suggest answers drawn only from the person's attached statements, each citing the statements it relies on; accepted answers go into every prompt, and the profile shows them under Where they stand.
  • Create a persona from a speaker. An executive's detail page in Insights offers Create persona from this speaker, pre-loaded with their statements.

See Personas.

Sharper persona grounding​

  • Tuning notes take effect. Notes in a persona's Tuning Notes section are now part of the persona's instructions for chat, reactions and content. A Marketplace persona carries the notes left on its catalog original. Switch a note off to stop applying it.
  • Recent writing on the topic at hand. When a persona is linked to its own sources, the material it reads is now chosen for relevance to what you are asking about, labelled as that person's recent writing on the subject.
  • Full articles, not teasers. When a news feed publishes only a summary, Kumkuat fetches the full article where the publisher allows it, and upgrades teasers filed earlier.
  • Cleaner excerpts. Reprint notices, copyright footers and subscription offers are stripped from news excerpts, so they are no longer read as the writer's own words.

More accurate speaker attribution​

  • No invented speakers. When a document has nothing to say on a theme, signal extraction no longer keeps a quote attributed to a person the document never names. A statement whose named speaker does not appear in the text is kept with its speaker marked Unknown rather than by-lined to someone who did not say it — for example, a partner press release that mentions a company but never quotes its CEO.

Home page​

  • Saved panels in the quick scorecard. The Scorecard card now leads with Your panels, like the Chat and Report cards. One click picks every member of a panel; a second click clears them, and un-picking one member by hand keeps the rest.

Audience pages on phones​

  • Audience Library. On a phone the category rail becomes a sideways-scrolling strip of chips above the list, so it no longer covers the audience rows as you scroll.
  • Audience Dashboard. Pinned audience cards and category tabs fit a phone screen: swipe through them, with the next card peeking in, and long category names wrap instead of being cut off.

Privacy, terms and Trust & Security​

  • Privacy policy and terms of service are published at app.kumkuat.ai/privacy and app.kumkuat.ai/terms, and linked from the sign-up and sign-in pages.
  • Trust & Security on this site gathers how Kumkuat handles your data: an overview, the sub-processors with a dated change log, a security & compliance FAQ with one link per questionnaire topic, and a control map against SOC 2 and ISO 27001.
  • Reporting a vulnerability. app.kumkuat.ai/.well-known/security.txt names security@kumkuat.ai as the contact for security reports.
  • Passwords are at least 8 characters for new accounts that sign up with email and password. Existing passwords keep working.
  • Private stays private from Kumkuat staff too. A private project or audience panel is visible only to its owner and the people it is shared with; Kumkuat support staff see what a member of your workspace would see.

For developers​

  • Developer portal reference. On developers.kumkuat.ai the reference is grouped by resource, with a page per operation and a shared errors section.
  • Canonical API paths. Endpoints now follow one pattern — /api/v1/<resource>, /{id}, actions as POST on their resource, no workspace id in the path. New homes include /api/v1/audiences (with /categories, /grades, /refresh, /{category}/summary and /reactions), /api/v1/personas/{persona_id}/summary, /api/v1/signals/templates, /signals/themes and /signals/filters, /api/v1/auto-tags/aliases, /api/v1/workspace/pull-through-priorities, /api/v1/documents/extract-text, /api/v1/executives/{executive_name}/voice (now readable with GET) and /api/v1/communications/filters.
  • One ingest endpoint. POST /api/v1/documents/ingest detects the content type by default, and YouTube links take the transcript path.
  • Nothing breaks. The older paths keep working with the same authentication, body and response; the reference marks them deprecated and names the replacement.
  • A clearer reference. The reference is curated to the operations meant for integrators — internal, sign-in and legacy routes are no longer listed. Operations now show their request and response models and query parameters under plain-language section names, and persona endpoints (and the matching MCP tools) no longer ask for a profile_id path parameter that does not exist. GET /api/v1/jobs/{job_id} is documented for following background work.
  • Agents can name a persona. POST /api/v1/personas/interact accepts a persona's name in place of its id when the name fits exactly one persona in the workspace, accents and all. If it fits none or several, the API answers 400 with the list of personas to choose from instead of a 500. This also helps assistants calling Kumkuat through MCP.
  • New guides. The Kumkuat API guide and Using AI assistants with the Kumkuat API join four how-to guides on rebuilding app features over the API: audience reactions, media pitching, message creation and the Audience Library.