Production / Service evidence
Monitoring & Observability
Instrument the boundary your users cross. Start from the symptom you need to catch, then keep only the signals that drive an action.
Workbench
Start from the symptom, not the metric
Instrument what a user could complain about. Each symptom below maps to one measurement you can take at your own boundary, a threshold worth reacting to, and the first action.
- Measure
- Share of retrievals returning zero results, plus a fixed set of retrieval questions with known answers.
- React when
- Empty-result share moves outside its normal band, or the known-answer set starts failing.
- First action
- Investigate what was written and how it is queried before suspecting availability.
Traceability
Correlate one request end to end
Emit one event per memory call with a stable shape. Low-cardinality fields keep it queryable; correlation identifiers make a single user report reproducible.
operation memory.query // fixed set of names outcome success | error | timeout status 429 // as returned by the API duration_ms 412 // measured around your call attempts 2 // including retries release web-2026.08.1 request_id X-Request-ID // echoed on the response trace_id X-Trace-ID // pass yours in to reuse it
Signal hygiene
Alert discipline
- Page only on sustained user impact; route trends to a review queue instead.
- Every alert names an owner and the first action, or it gets deleted.
- Require a minimum sample size so low traffic cannot trip a percentage alert.
- Run synthetic checks with disposable data, and never with a real end user’s identity.
Was this page helpful?