Memory#

Losing you isn't the end. An APS AI knows why it lost you, what hid you and where you were heading, keeps a diary of every encounter, and learns how you play.

What it knows the moment it loses you#

On AI Lost (Target, Last Known, Predicted, Last Cover Actor) is the start of every good search. The guard's reaction is your Blueprint; the knowledge comes from APS:

Why it lost you#

What hid you#

When it loses you from Detected or Tracked, APS traces from the guard's eyes to where you are before the event fires, and records the first thing in the way: Last Cover Actor on On AI Lost, or Get Last Seen Cover Actor. It is cleared on every loss, so it is always about this loss. Look behind it.

Where you were heading#

With Enable Prediction on (it's off by default: a slow patrolling guard that predicts feels psychic), Predicted is where your last run would have taken you, Prediction Horizon ahead. The guard measures your velocity from its own samples, so a fast run reads as fast. See Where it thinks you are for the details.

Short-term and long-term memory#

• Short term: after a loss, confidence fades (Lost → Remembered → Expired). Min Time in Lost makes it search at least that long before giving up; Decay Curve (or Default Decay Exponent) shapes the fade.
• Long term: Long Term Memory Strength is a second number on the belief, 0 to 1 (read it with Get Belief Data). It builds by 0.02 a second while the AI senses you, so it takes long contact to fill, and fades by about 1% a second after (or as Long Term Decay Curve says). APS itself doesn't act on it: read it to keep a guard twitchy for a minute after an encounter.
• Coming back: while you are Lost or Remembered, the first time a sense picks you up again adds Memory Refresh Bonus to confidence, once per loss, so it locks on faster the second time. A faint glimpse stays faint: one bonus, however fast the senses run.

A diary of every encounter#

APS keeps the last 5 encounters with each target. Read them with Get Recent Episodes or Get Last Episode; clear them with Clear Episodes.

An encounter is written when the AI loses you straight from Detected or Tracked. With the default Loss settings only Vision and Echolocation do that (they have direct cut). A loss that fades down through Suspected first, such as a sound or a scent going quiet, writes no encounter.

The payoff: the last three times I lost this player, they were Occluded within 2 m of the same pillar, so search there first.

A memory store your game writes to#

APS gives each AI somewhere to put a thought. What goes in and what it means are yours: APS doesn't decide that clothing matters, or vehicles, or which hand you swung with.

A memory is a tag on a subject (an actor, or a place), plus optional data:

Count and Age are the ones that make behaviour interesting: he has done this four times and that was a long time ago are different beliefs from he did this.

There is no "on recall" event, on purpose: deciding when to remember is the interesting part of your AI. Call Recall where your logic wants to know: on detect, in a decorator, when choosing a line of dialogue.

On AI Detect (Target)
  → Make Actor Subject (Target) → S
  → Remember (S, "SeenAtScene", Place: Target's location)
  → Branch: holding a weapon? → Remember (S, "Armed")

On AI Damaged (Instigator, Amount)
  → Make Actor Subject (Instigator) → S
  → Remember (S, "AttackedMe", Number: Amount)

Recall Tag (S, "AttackedMe") → Memory
  → Count ≥ 3   → not a suspect any more: a threat
  → Age > 120   → a while ago: challenge rather than open fire

It learns how you play#

An AI that notices you always crouch, always break away behind it, always hide behind the same crate, and starts checking there first. This is the player model, and it is automatic (Enable Player Behavior Model, on by default).

It is fair by design: it only records what the AI could perceive. It never reads your input. It watches your speed and stance while it can see you, and records where it lost you. If it never saw you crouch, it doesn't know you crouch.

Read it with Get Player Behavior Model (Target):

Always check Has Enough Data before adapting: it stops the AI concluding anything from one encounter.

• It observes while you are Detected or Tracked. An encounter closes, and the ratios update, only when Tracked turns into Lost. One that peaks at Detected and fades rolls into the next; a cautious player who never lets it reach Tracked teaches it nothing.
• So mid-fight you read the previous encounters: established history, which is what you want to act on.
• Reset Player Behavior Model (Target) forgets it all: a story beat, a difficulty reset, an accessibility option. Like memory, it lasts for the session.

Adapt behaviour, not numbers. Secretly sharpening the AI's eyes against a sneaky player feels like cheating, even when players can't name it. Change where it looks instead: patrol routes, facing, search order, how long it lingers. And let the player see it: a line of dialogue ("not this time"), a changed patrol. Adaptation nobody can perceive is just the game getting harder for no reason.

APS and Unreal's AI Perception#