Memory metrics

Memory metrics are served on port 9090. Metric names below include the Prometheus _total and _seconds suffixes where emitted by the client library.

API and retrieval

Metric	Labels	Meaning
omnia_memory_api_requests_total	method, route, status_code	HTTP requests
omnia_memory_api_request_duration_seconds	method, route, status_code	HTTP latency
omnia_memory_api_memories_saved_total	—	Memories saved
omnia_memory_api_memories_retrieved_total	strategy	Memories retrieved
omnia_memory_api_retrieval_duration_seconds	strategy	Retrieval latency
omnia_memory_retrievals_total	tier	Memories served by multi-tier retrieval, by tier
omnia_memory_retrieval_denied_total	—	Results removed by the denyCEL access filter
omnia_memory_rerank_fallback_total	reason	Reranking fell back to the fused order
omnia_memory_writes_suppressed_total	category, reason	Writes dropped before storage (user-opted-out, privacy-unavailable, consent-not-granted, unknown-category)
omnia_memory_accessed_updates_total	—	accessed_at updates from retrieval

Embeddings and ingestion

Metric	Labels	Meaning
omnia_memory_embedding_coverage	workspace	Fraction of live entities whose latest active observation has an embedding; the embedding’s model is not checked
omnia_memory_reembed_backlog	workspace	Active rows awaiting current-model embeddings
omnia_memory_embed_requests_total	outcome	Embedding calls
omnia_memory_embed_duration_seconds	—	Embedding latency
omnia_memory_reembed_rows_total	model	Rows the backfill worker has embedded
omnia_memory_reembed_pass_duration_seconds	model	Wall time of one backfill pass
omnia_memory_reembed_concurrency	model	Configured backfill concurrency
omnia_embedding_tokens_total	workspace, model, provider, source	Tokens charged to embedding usage
omnia_memory_ingest_documents_total	strategy, outcome	Documents ingested
omnia_memory_ingest_items_total	strategy	Items (chunks or summaries) persisted per ingest
omnia_memory_ingest_duration_seconds	strategy	Ingest latency

Coverage below 0.7 causes the Memory Galaxy to use a lexical basis. A growing re-embed backlog means newly ingested or stale rows may not participate in semantic retrieval.

The rate of omnia_memory_reembed_rows_total measures backfill throughput. The slope of omnia_memory_reembed_backlog is not throughput: the backlog falls by rows embedded minus rows newly written, so an ingest in progress holds it flat while the worker runs at full speed.

omnia_memory_reembed_concurrency reports the configured value, not a measured one, so a change in throughput can be attributed to a change in configuration. See Operate embeddings for measuring what it should be.

Retention, cache, and workers

Metric	Labels	Meaning
omnia_memory_retention_soft_deleted_total	tier, branch	Soft deletions
omnia_memory_retention_hard_deleted_total	—	Hard deletions after grace period
omnia_memory_retention_run_duration_seconds	outcome	Retention pass latency
omnia_memory_retention_run_errors_total	tier, branch	Retention branch failures
omnia_memory_worker_running	name	Worker liveness (1 when running)
omnia_memory_redis_up	—	Last Redis operation succeeded
omnia_memory_redis_dependency_errors_total	op	Redis dependency failures
omnia_memory_cache_lookups_total	op, result	Cache lookups
omnia_memory_cache_version_bumps_total	—	Scope-version bumps issued by writes

Enterprise workers also emit omnia_memory_projection_* metrics: workspace and policy labels, plus status and basis on the render counter.

Consolidation

Metric	Labels	Meaning
omnia_memory_consolidation_passes_total	workspace, service_group, policy, function, status	Consolidation passes. status: ok, empty, prefilter_error, function_error, apply_error
omnia_memory_consolidation_pass_duration_seconds	workspace, service_group, policy, function	Duration of one consolidation pass (per axis)
omnia_memory_consolidation_actions_total	workspace, service_group, policy, function, action, outcome, target_tier	Actions emitted. outcome: applied or rejected_<reason>
omnia_memory_consolidation_function_call_duration_seconds	workspace, service_group, policy, function	Duration of the HTTP call to a consolidation function

service_group scopes a sample to one of the workspace’s service groups, and distinguishes them when a workspace runs more than one. A worker started without --service-group emits an empty value.