Decay

Evidence is only as good as it is recent. Someone who looked at pricing a month ago is not someone who looked at pricing today. oracular models that with exponential decay governed by each intent's halfLife.

The rule

Before any update or reading, the intent's evidence is decayed to the current time. With Δt the time since the intent was last updated:

d  = 2^(−Δt / halfLife)
α  = α₀ + (α − α₀) · d
β  = β₀ + (β − β₀) · d

Each per-signal contribution is scaled by the same d. Only the excess over the prior decays, so with no events p returns to prior and evidence (α + β) returns to priorStrength, never to zero.

After six full-weight pricing.viewed signals in the quickstart config (α = 18.5, β = 9.5, half-life one week):

time since	p	lower	evidence
now	0.661	0.548	28.0
1 day	0.639	0.521	26.3
1 week	0.500	0.357	19.0
2 weeks	0.345	0.190	14.5
4 weeks	0.146	0.016	11.1
8 weeks	0.057	0.000	10.1

Decay is lazy

Decay is applied when an intent is touched: by observe for an event that is one of its signals or its outcome, and, without storing anything, by predict, shouldReveal and explain for the reading. An intent that nothing has touched for a month keeps its old numbers in the state and its old updatedAt; the next touch applies a month of decay in one step. The result is identical to decaying continuously, and the state stays cheap to update.

Evidence has a ceiling

Because evidence decays while it accumulates, a signal that arrives regularly settles at a steady level rather than growing forever. A weight w arriving every τ contributes at most about

w / (1 − 2^(−τ / halfLife))

in the long run. With the quickstart config, one pricing.viewed (+3) every three days with a one-week half-life settles near α ≈ 12, p ≈ 0.56, lower ≈ 0.44: it never fires at the default threshold. The same signal daily fires after seven (each day is a new session, so each adds its full weight), while repeating it within one session adds less each time. Half-life, weights and threshold have to be chosen together; the Tuning guide walks through it.

Choosing a half-life

Ask how long a signal should keep mattering:

• Stuck in onboarding: minutes. Being stuck is a state of the current session.
• Ready to upgrade: a week or two. Pricing viewed last Tuesday still counts.
• About to churn: weeks. Churn builds slowly.

Clocks

at is a timestamp and every duration is in the same unit. If a clock goes backwards (a device clock adjustment, events arriving out of order), the engine applies no decay and does not move its anchor back. Nothing throws and nothing becomes negative.

Learned data decays separately

Hit and miss counts, and the episode counts, are not evidence about now; they are knowledge about this user. They do not decay unless you set learning.halfLife, which gives them their own, usually much longer, half-life. See Episodes & learning.

Reference

• IntentConfig.halfLife, LearningConfig.halfLife, IntentState.updatedAt.