Observability

GET /metrics reports the silent failures fixed by the SQLite write queue, the CLI encoding guard, and empty auto-capture hints. Install the server extra, which includes prometheus-client:

pip install "neural-memory[server]"
nmem serve

nmem serve binds 127.0.0.1 by default. Scrape http://127.0.0.1:8000/metrics. The path is public, like /health, so a local scraper works while an API key is set. The body is counters and a latency histogram. It has no brain names and no memory text. A process without prometheus-client still answers the route with a one-line comment, and every recorder is a no-op.

Series

Name	Labels	When it moves
neural_memory_sqlite_write_latency_seconds	dialect="sqlite"	Every SQLite write. The sample includes queue wait and execution.
neural_memory_sqlite_write_collision_total	dialect="sqlite"	The driver reports SQL statements in progress. The single-writer queue keeps this at zero.
neural_memory_cli_encoding_damage_rejected_total	command, damage_type	nmem remember refuses damaged text. damage_type is question_mark_flood, unicode_escape, or replacement_char. --force stores the text and does not increment.
neural_memory_auto_capture_empty_total	reason, language	Detection found nothing. reason is too_short, cjk_unsupported, or no_trigger. language is zh, ja, ko, vi, or en.

command is limited to remember, forget, and list. A value outside a fixed set is recorded as other. The client also emits a _created gauge per series. Chart the names in the table.

Forbidden-pattern checks stay in scripts/pre_ship.py. That run fails the build in CI. It is not a runtime series.

Grafana

Import docs/grafana/neural-memory-silent-failures.json and point it at the Prometheus datasource. Load docs/grafana/neural-memory-alerts.yml as a Prometheus rule file.

The collision counter is cumulative, so a single historical event would leave neural_memory_sqlite_write_collision_total > 0 forever. The shipped rule fires when the counter increases inside five minutes:

increase(neural_memory_sqlite_write_collision_total[5m]) > 0

Severity is warning. An increase means a commit collision reached the driver again.