Please use the form below to provide your feedback. Because your feedback is valuable to us, the information you submit in this form is recorded in our issue tracking system (JIRA), which is publicly available. You can track the status of your feedback using the ticket number displayed in the dialog once you submit the form.

Managed Agent Memory Metrics and Alerts Reference

  • reference
Reference for the metrics on the Managed Agent Memory monitoring dashboard and the alerts that Capella sends for a Managed Agent Memory server.

For how to view the dashboard and receive alerts, see Monitor Managed Agent Memory.

Metrics

The following table lists the Agent Memory server metric behind each card and chart on the monitoring dashboard. Rates and percentiles use a 5-minute window.

Section Card or chart Metric Calculation

API

Requests total

agentmemory_requests_total

Rate per second, summed across endpoints.

API

API errors total

agentmemory_api_errors_total

Rate per second, summed across status codes.

API

Blocks added

agentmemory_blocks_added_total

Rate per second, summed across block types.

API

Request Duration

agentmemory_request_duration_seconds

99th percentile, for each endpoint.

API

Tokens per Memory Block

agentmemory_tokens_per_block

95th percentile, for each type.

Models & search

Embedding failures

agentmemory_embedding_failures_total

Rate per second.

Models & search

Context failures

agentmemory_context_generation_failures_total

Rate per second.

Models & search

Embedding Generation Duration

agentmemory_embedding_generation_duration_seconds

99th percentile, for each model.

Models & search

Context Generation Duration

agentmemory_context_generation_duration_seconds

99th percentile, for each model.

Ingestion queue

Queue size

agentmemory_async_batch_queue_size

Sum of queued items across nodes. Items in progress are not included.

Ingestion queue

Queue capacity

agentmemory_config_queue_max_size

Sum across nodes. Equals your Queue max size setting multiplied by the number of nodes.

Ingestion queue

Extraction success rate

agentmemory_async_batch_process_outcomes_total

Successful outcomes as a share of all outcomes.

Ingestion queue

Extraction failures

agentmemory_async_batch_process_outcomes_total

Rate per second of failed outcomes.

Ingestion queue

Queue Wait Time

agentmemory_queue_wait_time_seconds

99th percentile.

Ingestion queue

Extraction Duration

agentmemory_async_batch_process_duration_seconds

99th percentile.

Rate budget

Tokens available

agentmemory_rate_budget_available_tokens

Sum across nodes.

Rate budget

Max token budget

agentmemory_rate_budget_max_tokens

Maximum across nodes. Equals each node’s share of your Tokens per minute setting.

Rate budget

Requests available

agentmemory_rate_budget_available_requests

Sum across nodes.

Rate budget

Max request budget

agentmemory_rate_budget_max_requests

Maximum across nodes. Equals each node’s share of your Requests per minute setting.

Health

Overall health

agentmemory_health_check_status

Lowest value of the dispatcher, async loop, and database components. 1 means healthy.

Health

Dispatcher health

agentmemory_health_check_status

Dispatcher component. 1 means healthy.

Health

Async loop health

agentmemory_health_check_status

Async loop component. 1 means healthy.

Health

Database health

agentmemory_health_check_status

Database component. 1 means healthy.

Health

Model health

agentmemory_health_check_status

Lowest value of the embedding model and LLM components. 1 means healthy.

Health

Live nodes

agentmemory_cluster_live_node_count

Maximum across nodes.

Alerts

Capella sends the following alerts for each Agent Memory server. Capella checks each alert on each node, except the 2 retrieval latency alerts. You cannot change their conditions.

Alert Severity Fires when

Agent Memory Read/Write Error Rate

Critical

More than 2% of requests return a service error (5xx) over 5 minutes, with at least 10 service errors.

Agent Memory Model Provider Error Rate

Critical

More than 5% of calls to your embedding model or LLM fail over 5 minutes, with at least 10 errors. Check the status of your model provider and your OpenAI integrations.

Agent Memory Queue Depth

Critical

The extraction queue, including items in progress, is more than 90% full. When the queue is full, requests to add memory return HTTP 429 with a Retry-After header.

Agent Memory Model/DB Unreachable

Critical

Your server cannot reach your models or your cluster for more than 5 consecutive health checks.

Agent Memory Key Revoked

Critical

The OpenAI API key for your embedding model or LLM is deleted or revoked. Your server needs a valid key.

Agent Memory KV retrieval latency has degraded and requires investigation.

Critical

The 99th percentile latency of key-value reads over 5 minutes is more than twice the 99th percentile for the previous hour, with at least 20 samples.

Agent Memory semantic search retrieval latency has degraded and requires investigation.

Critical

The 99th percentile latency of semantic search over 5 minutes is more than twice the 99th percentile for the previous hour, with at least 20 samples.

Agent Memory Custom Prompt Fallback

Warning

Within the last 15 minutes, Agent Memory extracted memory with the default prompt because it could not load a custom prompt. For more information, see Customize Extraction and Summarization Prompts.

Agent Memory Storage Capacity Warning

Warning

A node has used 80% or more of its disk storage. If the disk fills, requests to add memory fail.

See Also