io.github.PeterAtPondo/kit-memory-demo

Kit memory demo · hosted MCP servers

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.

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.

Quickstart

Claude Code:

claude mcp add --transport http kit-papers https://demo-mcp.kit-project.com/mcp \
  --header "Authorization: Bearer kbr_live_Kh9eO_ZfLldveFf1KXJEEvepfOpdgRqf7KoDDU3rB18"
claude mcp add --transport http kit-mews https://mews-mcp.kit-project.com/mcp \
  --header "Authorization: Bearer kbr_live_UN8-2LUlKmP2uxWQB5vpn-fzkJIuRb5zmq-apQeTU7U"
claude mcp add --transport http kit-case-file https://real-mcp.kit-project.com/mcp \
  --header "Authorization: Bearer kbr_live_ISWsqW39NkqT659FEkQjv0OjJDmyD4O3wx3wcCfA03c"

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.

No MCP client? No problem

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.

Scoped access, demonstrated

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:

kbr_live_FoAauFWl7_IoDswh8rJrFvCw2qnblQTOF8DuSBkbj7A

What these servers are, honestly

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.

License

MIT (code and examples in this repo). The demo worlds themselves are licensed with the benchmark: see kit-memory-eval's DATA-LICENSE.