Prometheus¶

Membrick exposes a Prometheus text endpoint at GET /metrics. Prometheus can scrape it to collect aggregate query and agent activity metrics.

Membrick currently supports Prometheus exposition format. It does not currently export OpenTelemetry OTLP metrics, traces, or logs.

Start the Membrick server¶

Start the server for the project you want to monitor. For a local Prometheus instance, bind Membrick to localhost:

cargo run --release --bin server -- \
  --settings /path/to/project/.membrick/server.toml

Replace /path/to/project with your project directory. Put the token in the server_token setting. For a single-user local setup, you can instead use --allow-unauthenticated; keep the server bound to localhost.

Configure Prometheus¶

Add a scrape job to prometheus.yml. When Membrick uses a server token, put the same token in a file readable by Prometheus and set bearer_token_file:

scrape_configs:
  - job_name: membrick
    scrape_interval: 15s
    metrics_path: /metrics
    static_configs:
      - targets: ["127.0.0.1:8080"]
    bearer_token_file: /path/to/membrick-server-token

For an unauthenticated localhost-only server, omit bearer_token_file. Reload Prometheus after changing its configuration, then check Status → Targets in the Prometheus UI. The membrick target should show as UP.

You can also inspect the endpoint directly:

curl -H "Authorization: Bearer replace-with-your-server-token" \
  http://127.0.0.1:8080/metrics

Metrics¶

Search activity¶

Metric	Type	Meaning
`membrick_query_requests_window`	Gauge	Search requests in the retained telemetry window
`membrick_query_errors_window`	Gauge	Search errors in that window
`membrick_query_empty_results_window`	Gauge	Searches with no results in that window
`membrick_query_requests_per_second`	Gauge	Recent search request rate
`membrick_query_error_rate`	Gauge	Fraction of searches that returned errors
`membrick_query_empty_result_rate`	Gauge	Fraction of searches with no results
`membrick_query_latency_ms{quantile="0.5"}`	Gauge	Estimated median search latency in milliseconds
`membrick_query_latency_ms{quantile="0.95"}`	Gauge	Estimated p95 search latency in milliseconds

These search values describe the rolling telemetry window (up to ten minutes), so they can go down as older buckets expire. The latency quantiles are estimates from Membrick's existing latency buckets, not Prometheus histograms.

Agent activity¶

Provider metrics use the bounded provider label values claude, codex, gemini-cli, agy, openclaw, and hermes.

Metric	Type	Meaning
`membrick_provider_compiled{provider}`	Gauge	Whether this server build includes the integration (1 or 0)
`membrick_provider_observed{provider}`	Gauge	Whether any lifecycle event has been received (1 or 0)
`membrick_provider_events_total{provider}`	Counter	Lifecycle events recorded for the provider
`membrick_provider_sessions_started_total{provider}`	Counter	Sessions started
`membrick_provider_sessions_ended_total{provider}`	Counter	Sessions ended
`membrick_provider_retrieval_events_total{provider}`	Counter	Events classified as retrieval
`membrick_provider_capture_events_total{provider}`	Counter	Events classified as capture
`membrick_provider_sessions_active{provider}`	Gauge	Active sessions observed by Membrick
`membrick_provider_last_seen_timestamp_seconds{provider}`	Gauge	Unix time of the last event, or zero if none was received
`membrick_provider_recent_events{provider,category}`	Gauge	Events by category in the retained provider ledger
`membrick_provider_token_usage_recent{provider,kind}`	Gauge	Token counts in retained events that reported usage

category is one of session, retrieval, compaction, capture, or lifecycle. It groups event kinds for monitoring; it does not expose each session or tool event separately. Recent event and token gauges summarize the provider ledger, which retains at most 500 events per provider. Token values are only available when the agent reports them.

Memory index¶

Metric	Type	Meaning
`membrick_index_source_documents`	Gauge	Source documents in the project index
`membrick_index_records`	Gauge	Indexed memory records
`membrick_index_dirty`	Gauge	Whether there are unpublished index changes (1 or 0)
`membrick_index_store_revision`	Gauge	Current mutable store revision
`membrick_index_snapshot_revision`	Gauge	Published search snapshot revision
`membrick_index_revision_lag`	Gauge	Store revision minus published snapshot revision
`membrick_index_segments`	Gauge	Segments in the published search snapshot

For example, graph recent p95 latency and active sessions in Prometheus with:

membrick_query_latency_ms{quantile="0.95"}
membrick_provider_sessions_active

Prometheus receives aggregate metrics only. It does not receive query text, memory contents, session identifiers, prompts, tool arguments, or the dashboard's individual event timeline. See Telemetry details for the retained telemetry data and Observability for the dashboard.

Network access¶

The metrics endpoint uses the server's configured authentication. Keep Membrick on localhost unless you have secured the network path and token distribution. See the HTTP server guide for authentication and bind-address behavior.