Skip to main content
An Engram adapter translates a memory provider’s observable API results into Memory Telemetry v2. It should be structurally thin and evidence-preserving.

​
Required behavior

  1. Wrap provider calls without changing their return values or thrown errors.
  2. Emit only inside an active EngramClient.withTurn scope.
  3. Preserve concrete memory IDs, ranks, scores, provider/store identity, and useful metadata.
  4. Label translated provider responses mapped or observed according to the integration boundary and include a source path.
  5. Report an instrumentation gap instead of fabricating an operation when IDs or outcomes are unavailable.
  6. Keep retrieval separate from context loading. The application must call turn.load after prompt construction.
  7. Remain fail-open unless the SDK was explicitly configured with strict: true.

​
Tests

Include deterministic coverage for successful operations, empty results, malformed/partial provider output, asynchronous acknowledgements, provider errors, and calls outside an active turn. Do not require live credentials in normal CI.

​
Review checklist

  • Does every emitted operation have evidence provenance?
  • Can an asynchronous write be mistaken for a completed store?
  • Can a retrieved record be mistaken for loaded context?
  • Are provider errors and return values preserved?
  • Are unknown response shapes surfaced rather than silently normalized?
  • Does the README document unsupported operations and provider-version assumptions?