Tune confidence decay

Memories in Brain carry a confidence score. A per-shard background worker decays that score over time so old, unreinforced memories drift down the ranking without being deleted. Decay is a deploy-time worker, not a per-request or per-namespace runtime setting — you tune it in Brain's config and restart.

The knob

Decay cadence lives in the [workers] section of config.toml (baked into the image at /etc/brain/config.toml). The default runs hourly:

[workers]
decay_interval_sec = 3600   # run the decay pass every hour

Lower it to decay more aggressively for a fast-moving deployment:

[workers]
decay_interval_sec = 300    # every 5 minutes

Any field can be overridden at runtime with the generic BRAIN__SECTION__FIELD env form — no bespoke variables:

BRAIN__WORKERS__DECAY_INTERVAL_SEC=300

The worker runs on every shard. The new cadence takes effect on the next server start.

Verify

Decay lowers the confidence of unreinforced memories over successive passes. Recall a known cue and watch the field move between runs:

const answer = await brain.recall({ query: 'last quarter revenue', max_results: 10 });
for (const m of answer.memories) {
  console.log(m.memory_id, m.confidence, m.salience);
}

answer = brain.recall('last quarter revenue', max_results=10)
for m in answer.memories:
    print(m.memory_id, m.confidence, m.salience)

Re-encoding the same fact reinforces it and resets its decay — a memory that keeps coming up stays confident; one nothing refers to fades.