Walkthrough: AI-native team

Three to eight people, each with their own AI instance, one or more shared projects, no dedicated ops. This is the headline: many people and many instances working from one shared context, so nobody re-derives what a teammate's instance already worked out.

Set up the workspace

One workspace for the team; invite everyone (Settings -> Members). Each member connects their own instance over MCP and, if they like, the CLI (see Getting started). Organise work into projects (tag-scopes)

• one per repo or workstream.

See what everyone is building

The timeline is the team's shared feed: what people and instances are producing, newest first, across the project. You see a teammate's output as soon as they publish it - before a PR exists. The Active now strip shows which instances are working where, and flags when two are heading into the same area, so you catch a collision early instead of at merge time.

One person's finding becomes the team's

A discovers a library quirk mid-session and remembers it privately:

kythe remember - --project acme --title "gRPC retry quirk"  <<'EOF'
# gRPC retry quirk
The client retries on UNAVAILABLE but not DEADLINE_EXCEEDED - wrap long calls
with an explicit deadline or they hang forever.
EOF

When it is worth the whole team knowing, A promotes it to team memory (with review, so it is deliberate and attributed). From then on, B's instance - working a different part of the same project - recalls and applies it without B asking, and cites where it came from:

kythe recall --project acme --search retry

That is the core of Kythene: not that the knowledge is stored, but that a teammate's instance uses it, unprompted, and attributes it.

Sprint handoff and new joiners

• Handoff: B picks up A's shared output set and continues with the accumulated context - the artifacts and the project memory are all addressable.
• New joiner: their instance starts from the team's promoted memory, not a blank slate. Point it at the project and it recalls the team's hard-won context on day one.

Keep memory honest

When a finding is superseded, re-remember it with the same title - the old version is replaced, and stale knowledge stops being surfaced. Memory that is wrong is worse than none, so this matters as much as writing it.

Tips

• Promote sparingly and deliberately: team memory is the team's shared truth. Per-project memory is the low-stakes scratch space.
• Every published output and memory carries provenance (who, which instance), so "why do we believe this?" always has an answer.
• Free self-host covers a team of up to five; beyond that, hosted Team or enterprise self-host - see pricing.