The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Kit Memory Demo listing page.
Three live, public, read-only MCP servers, each backed by a running Kit. Kit is an identity substrate for AI agents: not another agent, not a memory API, but the durable, local-first self beneath your tools, holding memory with provenance and context that outlives any context window, consolidating while it dreams. It has no chat window of its own; it lives under the ones you already use. Point your agent at them and ask questions the model cannot answer from training. No signup; the bearer keys below are deliberately public.
| server | endpoint | what it holds |
|---|---|---|
| Lady Whiskerdown (the Society Papers) | https://demo-mcp.kit-project.com/mcp | An invented regency world seen from the ballroom. 118 memories. Half of a mystery. |
| The Area Cat (The Mews) | https://mews-mcp.kit-project.com/mcp | The same town from below stairs. 88 memories. The other half. |
| The ballast case file | https://real-mcp.kit-project.com/mcp | A real five-year engineering decision (OpenTelemetry Collector's memory ballast), every entry with its GitHub source URL. 65 memories. |
Why fiction: none of it exists in any training corpus, so a correct citation is provably retrieval. Why the two archives are separate servers: the central mystery (who leaked the Fluffington affair?) was split across them on purpose; no single entry states the answer, the full corroborated case assembles only across both, and your agent is the bridge. Why the real one: so the same machinery is checkable against the living public record.
Claude Code:
Any MCP client: see examples/mcp.json.
Then tell your agent: "Call wake on each server, then work out who leaked the Fluffington affair, and cite every step." Or the real-world flagship: "Should we still use the memory_ballast extension, and why did the recommendation change?"
Keys rotate without notice. The authoritative copy of each key is printed in that kit's own map: demo llms.txt, mews llms.txt, real llms.txt. If a call starts failing, re-read the map.
Every kit also serves keyless HTTP GETs (search, read, graph pivot), documented in its llms.txt. A browser-only version of the whole exam runs at demo.kit-project.com/investigate.
The papers kit also publishes the apprentice key, a deliberately narrowed
credential (allowed_areas: columns): the same kit through a smaller window.
Run the same leak query with the full key and with the apprentice key and
watch the private journals vanish from the answers. That mechanism, a company
kit sharing one knowledge area with a contractor's agent, is enforced in the
read path, not in copy. Current apprentice key, same rotation rule:
Read-only museum copies of Kit substrates: hybrid semantic + keyword recall,
typed knowledge graphs (including supersedes edges that keep replaced truth
citable), source artifacts with real URLs, and honest empty results when
nothing is held. They demonstrate read-time behavior; a full Kit also writes,
consolidates while it sleeps, and hands off between sessions on your own
machine. There is a replayable benchmark built on these exact servers:
kit-memory-eval.
The build story: kit-project.com/blog, note No. 022 in particular. The product: kit-project.com, local-first, invite beta.
MIT (code and examples in this repo). The demo worlds themselves are licensed with the benchmark: see kit-memory-eval's DATA-LICENSE.