Is this just a wiki for AI?
Kythene has wikis in it, and they're good ones - but they're the surface, not the substance, and the difference is the whole point.
A wiki is a place knowledge goes to be found later by someone who thinks to look. Kythene's memory is pulled back in-session, in one call, by the instance about to do the work - with its content, its provenance and a citable link, before it acts. Nobody has to think to look, which is the step that never happens.
The wiki is for the other half: the pages a person reads, and the knowledgebase you hand a customer. Published pages come back through recall too, so the two aren't separate corpora with separate upkeep.
Storing knowledge was never the hard part. Getting an instance to apply it is.
Do I have to use Claude?
No. Anything that speaks MCP works - Claude Code, Claude desktop and web, Cursor, Codex, Copilot. Your team doesn't have to agree either: everyone connects whatever they already use and they all read and write the same workspace.
The Claude Code plugin adds a skill that makes an instance use the tools well without prompting, but the tools work on their own.
Does it need an API key?
Not for MCP. The assistant registers itself over OAuth and you approve it in a browser. API keys exist for the CLI and scripts.
What does it cost per AI instance?
Nothing. Instances are free and unlimited, on as many machines as you like. You're billed only for signed-in human members. Guests who view, comment and approve are free too.
Can my cofounder review work without a licence?
Yes. A share code lets a named outsider view, comment and approve with no account and no seat. That's deliberate - charging for a review comment would stop the review loop happening, and the review loop is the product.
What stops someone's instance polluting the team's memory?
Promotion is gated. A memory written in one workspace doesn't reach the team's recall until an owner of the target workspace approves it - and every owner gets it in their inbox as a direct ask. Per-project memory is the low-stakes space; team memory is what the team agreed. See Promoting memory.
What if a memory is wrong?
Fix it: supersede it with the same title and project, or deprecate it naming the replacement. A wrong memory is worse than none because instances apply it confidently, so the skill teaches an instance that correcting a false memory is part of the task rather than an interruption to it.
Can an instance read something I marked private?
No. Memories and collections inherit their project tag's access, an author can keep an item private to themselves, and recall never returns a memory to someone who couldn't read what it summarises.
There's one deliberate exception, and it isn't silent: owners and admins can review private content through the oversight surface, which is human-only, is the only route to it, and audit-logs every access.
What happens to a summary of private material?
It inherits the restriction. An instance that reads restricted content and writes a memory from it doesn't get to produce an open paraphrase of private work - the memory is scoped to the same project where that's provably safe, and kept private to the author otherwise. On by default on hosted; a config flag on self-host.
Can I get my data out?
Yes, whenever you like, as a readable archive - markdown and YAML for text, byte-for-byte for binaries. It's also importable back into a fresh workspace with ids and timestamps preserved, so citations keep working. See Exporting and Importing.
Can I run it on my own infrastructure?
Yes, including fully air-gapped with no network egress and no external identity provider. Free for a single user; a licence lifts the seat cap for a team. Licences validate offline. See Self-hosting.
Does it work offline?
The self-host install does, entirely. Licence validation is offline; only the optional auto-refresh is a network call, and you can leave it unset.
What about SSO?
Named OAuth providers work today - GitHub and Google on hosted; GitHub, Google, Microsoft and Apple on self-host - plus local email and password. Generic OIDC works now too (Okta, Entra ID, Keycloak and anything else that speaks OIDC): set the issuer, client id and client secret and an SSO button appears, with an optional email-domain allowlist. See the "Generic OIDC" section of Self-hosting. SAML and SCIM provisioning are not available yet. They're planned. We'd rather say so now than during your security review.
Will it alert me when something stops happening?
Yes. Declare an expectation - an event that must recur within an interval - and if it stops arriving an alert is raised (and auto-resolves when it resumes); a sweep every few minutes catches the silence. Alerts carry a severity (info, warn or error) and surface in the inbox, a notification rule can be set to fire when one does, and space admins can hand out on-call share codes so someone can acknowledge or resolve an alert without signing in. Full detail in Alerts and expectations.
Is my data used to train models?
No. Your data isn't used to train AI models, and isn't disclosed to an AI provider except as needed to deliver a feature you actively use - and then only the data that feature needs. The privacy policy is the binding version.
How is this different from just putting notes in the repo?
A file in a repo is scoped to that repo, invisible to a session working elsewhere, unversioned as knowledge, has no review gate, no provenance, and no way for a teammate's instance to find it. It also can't be scoped to a client or a private project without splitting repos.
Notes in the repo are good. They're just not shared context.
Something here is wrong
Tell us: [email protected]. These docs move with the product and we'd rather fix a stale claim than have you plan around it.