View memory audits and statistics-AgentLoop(AgentLoop)-阿里云帮助中心

Check whether the data source is consumed as expected

Go to the overview page of the memory store and check the source type, filter conditions, consumption status, progress, and latest error.

Manual write: First, check the response of the write API. If write event logs are available, view the input in Write Events.
Agent trace: Check whether the current workspace contains traces that meet the conditions such as service name, agent name, number of steps, and start time.
Dataset: Check the bound dataset, version policy, and filter conditions, and make sure that the fields still conform to MemorySourceV1.

For trace or Dataset sources, you can use Pause Consumption and Resume Consumption to control subsequent reads. A page showing that the store is created or that no recent error exists does not by itself prove that a specific input has generated a long-term memory. Continue to inspect the extraction and storage results.

View audit details

Go to the Audit Details tab of the destination memory store.
Select the time range that you want to inspect and filter by Event Type.
Locate related events based on the input source, memory ID, recall event ID, or error message.

Audits help distinguish among the following situations: an input has entered processing, extraction produced no facts, storage added or updated a memory, a memory was skipped or is pending review, and a recall succeeded, failed, or returned an empty result.

Audit logs arrive asynchronously from business processing. If no record is available immediately after an operation, check the time range and filter conditions, and query again later.

Trace a recall

After you obtain a recall event ID on the Long-term Memory Retrieval page, you can open the corresponding audit record. You can also open the details of a recall from Statistics > Recent Recalls.

Focus on the following information:

Information	Purpose
Query text and business scope	Verify that the query targets the expected user, agent, application, or run.
Status, error code, and latency	Distinguish a normal empty result from an execution failure.
Number of results, ranks, and scores	Verify which memories were returned to the caller.
Returned content and revision	Verify which memory version was returned.
Drop stage	After related logging is enabled, identify the stage at which a candidate was discarded.

If query logging is set to Log HMAC Only, the original query text is not displayed. If an audit degradation or content truncation message appears, interpret the integrity of the record as indicated. Do not assume that no recall occurred merely because details are missing.

Interpret statistics

Metric or view	Meaning and usage
Recall count	The number of recall events after deduplication. Use it with the time range to observe request volume.
Non-empty recall rate	The percentage of successful recalls that return at least one result. Evaluate this metric together with result quality.
P95 latency	The 95th percentile of recall latency, which is used to identify high-latency cases.
Average result count	Observe the number of memories returned by each recall.
Extraction count	The number of input rows that were successfully extracted. This value is not the number of long-term memories that were ultimately added.
Total memories	View the current size of the memory store.
Recall funnel	Compare the numbers of candidates, candidates after threshold filtering, candidates after reranking, and final results.
Extraction funnel	Compare the numbers of pulled, queued, successfully extracted, written, skipped, and pending-review items.
Top recalled memories	Identify memories that are frequently returned.
Zero-result queries	Identify queries that repeatedly return no results, and inspect the scope, input coverage, and policies.
Recent recalls	Drill down from trends to individual requests.

Select the same time range before you compare trends with individual requests. The extraction funnel is aggregated by memory store. Do not assume that recall type, recall status, or similar filters also apply to extraction data.

The non-empty recall rate measures whether memories are returned. To evaluate whether memories improve agent responses, also check whether the application uses the returned content and assess the actual task results.

Common troubleshooting paths

Symptom	Recommended inspection order
No new memory is generated	Data availability at the source → Start position and filter conditions → Extraction result → Category or exclusion rules → Storage action
New information does not replace old information	Whether event append mode is used → Merge conditions and allowed actions → Whether manual edit protection is triggered
Memories exist but the query returns no results	Store and region → Consistency of business IDs → Additional filters → Query text and threshold-related records
The non-empty recall rate decreases	Whether new business traffic or query types were added → Business scope and memory coverage → Zero-result queries and individual request details
No data is displayed in the report	Time range and filters → Log resources associated with the workspace and access permissions → Whether audit records have arrived

To adjust extraction, storage, or audit logging, see Configure memory extraction and storage strategies.