This runs against the real memory API in an isolated demo scope. Nothing here touches real project data, and you can reset it at any time. Walk through the three steps below to see why rules are read deterministically while history is searched semantically.
step 1 — seed a contradiction
Two "rules" get written to the same subject, one right after the other — exactly what happens when an agent changes its mind mid-project.
step 2 — deterministic read (rules)
Call GET /rules for this scope. There is no ranking, no similarity score —
the superseded rule is gone, not just buried.
step 3 — semantic recall (episodes)
Rules are never fuzzy-searched — they're read exactly, which is what step 2 just showed.
/recall is for everything else: facts and history. Write a fact below, then search for it in
plain language different from how you wrote it. This is what agents use to pull up unstructured context that
a deterministic lookup can't index.
side by side
/rules — deterministic
Exact scope match. One current answer per subject. No score, because there's nothing to rank.
// call step 2 first
/recall — semantic
Vector similarity. Ranked list, contradictions included, because it's a search over raw history.
// call step 3 first
Everything above is running against a real, isolated scope
(—) created just for this browser session. Hit reset to start clean, or just leave —
demo scopes never appear in real project data.