Create a task, save memory, end the session — then resume as a brand-new agent and see exactly what it remembers and why. Every step runs real Continuity code on a real ledger created just for you. No account, no model calls, no mockups.
Step 2: type your own words as the memory — a scripted demo can't echo back text it has never seen.
There is no language model in this demo. Nothing here can hallucinate. You are testing the layer that survives when a model's session dies: an append-only, hash-chained SQLite ledger and the projections computed from it. That layer — not a model — is the product.
Why it can't be scripted: the memory in step 2 is your own words. It comes back in step 3 with the ledger sequence that stored it, and appears in the receipt under a hash chain computed from your specific ledger. Pre-recorded responses cannot contain text you just invented.
Verify offline: download your ledger above, then on your machine:
pip install continuity-mcp and
continuity-handoff --db continuity-playground.db --project-id playground --task-id checkout-refactor --out RECEIPT.md
— the shipped CLI reproduces the same receipt from the same bytes. The demo and the product are one code path:
continuity/handoff.py, continuity/projections.py, continuity/store.py, all open source, all covered by
public CI including adversarial tamper tests.
What this demo does not prove: that a specific LLM will use the compiled
context well. For that, the repo ships isolated, nonce-based persistence proofs you run with your own
Claude Code or Codex subscription (make claude-persistence-proof) — designed so the answer cannot
come from the model's native memory, only from Continuity retrieval.
Session B's panel is filled by compile_task_context — the same projection your
real agents call over MCP. The receipt is build_verified_handoff. Your ledger is deleted after
15 minutes.