AI Product Psychology

Peak-End Rule: Users Only Remember the Peak and the End

Nobel laureate Kahneman found that how people remember an experience is almost entirely set by the most intense moment (the peak) and the final moment (the end) — average level and duration barely matter. This rule answers a budget question directly: where in a session should you spend the most expensive model and the most optimization effort. This lesson has five experiments — play them one by one.

Play the plots first · Four fates of the same session

One 10-turn AI session, each turn scored 0 to 10. Open the four plots in turn and watch how average score (what engineering dashboards measure) and memory score (what users keep — roughly the average of peak and end, from Kahneman) diverge. Focus on plot ③: highest average of the four, lowest memory score.

Tap a plot — watch the two scores fight Seen 1 / 4
Tap all four before reading on — the verdict sits under the chart.
Average score (engineering view)
0.0
Arithmetic mean of ten turns
Memory score (what users keep)
0.0
(peak + end) ÷ 2
Unpack the source · That 1993 basin of cold water

The original evidence for the peak-end rule is a paper whose title is already a dare: Kahneman, Fredrickson, Schreiber & Redelmeier, 1993, in Psychological Science — When More Pain Is Preferred to Less: Adding a Better End. Subjects went through two cold-water trials, then answered one question: if you must do one again, which do you pick? Guess how most people chose.

Guess: which trial would most subjects redo? Tap a card to vote
Votes were cast after living through both trials. Tap one on gut feel.
Trial A · Short
14°C · 60 s
60 seconds total

One hand in 14°C water for a full 60 seconds, then done — towel off.

Trial B · Long
14°C · 60 s15°C · 30 s
90 seconds total

Same 60 seconds at 14°C, then 30 more seconds as the water quietly rose to 15°C: still unpleasant, just less biting.

Behind this is the split Kahneman keeps returning to in Thinking, Fast and Slow: the experiencing self lives every second (B objectively suffers 30 extra seconds), while the remembering self scores only the peak and end frames. Only the latter shows up to vote. In 1996 Redelmeier and Kahneman repeated the finding with real colonoscopy patients: procedures from 4 to 69 minutes, post-hoc pain ratings barely tracked duration, and tracked peak pain and the last three minutes tightly. The phenomenon has a name: duration neglect.

The experiencing self lives the day; the remembering self makes the decisions. Ratings, renewals, recommendations to friends — the remembering self signs them all. Your DAU curve measures the former; your NPS belongs to the latter. Waiting from lesson 2 works the same way: how long you waited is forgotten; how the wait ended is remembered.
Build the curve yourself · Fool the average score

Watching ready-made plots isn't enough — building one yourself is what sticks. Below are 10 nodes in a session; tap a node to cycle good (+2) / mid (0) / bad (−2), and the two scores on the right update live. The challenge: build a curve with average score below 0 and memory score maxed at +2.0. When you hit it, the badge lights up.

Experience-curve editor Challenge in progress
Hint: memory score only looks at two positions. Figure out which two and this takes ten seconds.
Average score
0.0
Arithmetic mean of ten nodes
Memory score
0.0
(peak + end) ÷ 2
Challenge cleared. The curve you just built would flash red on an engineering dashboard (negative average), yet users would tell colleagues it was a good experience (memory score +2.0). The reverse holds too: a session that looks all-green on the dashboard may already be fading from memory. Teams that monitor averages and users who store only peak and end are reading two different report cards.
Pull out ¥100 · Budget follows memory weight

If memory score only honors peak and end, inference budget should follow memory weight. Suppose a session has ¥100 of inference budget — drag three sliders across first interaction / middle / wrap-up (locked to sum to 100) and watch memory score and the verdict shift. One hidden rule up front: the middle has a semantic-cache floor at quality 5.0 — that's why it can be cheap. The cost track covered this when it taught tiered routing.

Model-budget allocator Memory score 6.7
Try three splits: even money, all-in on wrap-up, heavy at both ends and light in the middle. Badge lights up at memory score 7.4+.
¥33
¥34
¥33
First interactionMiddleWrap-up
First-interaction quality6.7
Middle quality (cache floor 5.0)6.8
Wrap-up quality6.7
Memory score
6.7
Wrap-up duel · What the last step should do

Budget done — now flow. Same Agent task: “generate a quarterly report and publish it externally.” Two wrap-up arrangements: A leaves external publish as the tenth-step finale; B moves validation to step one, persists intermediate artifacts as they go, and step ten only packages delivery plus a summary. Tap the version you think has the higher memory score.

Two wrap-up arrangements — pick the higher memory score Pick a side
Plan A
Agent · Generate quarterly report
✓Pull sales data source
✓Clean and aggregate
✓Generate charts and body copy
…steps 4 through 8 each pass…
✓Render 12-page report
✗Step 10 · Call external API to publish
Publish failed: third-party API 502. This session’s results were not saved. Please re-run the task.
Plan B
Agent · Generate quarterly report
1Validate permissions, quota, export format (up front)
✓Pull sales data source
Intermediate artifact persisted · report_draft_01.md
✓Clean, aggregate, generate charts and body
Intermediate artifact persisted · charts_q3.zip
✓Step 10 · Package delivery + summary
Report saved to your space: 12 pages, 2 spots marked for your review. Publish externally? (Optional — anytime.)
Sources and further reading: Cold-water experiment: Kahneman, Fredrickson, Schreiber & Redelmeier (1993), When More Pain Is Preferred to Less: Adding a Better End, Psychological Science; colonoscopy study: Redelmeier & Kahneman (1996), Pain; systematic account of remembering self, experiencing self, and duration neglect: Kahneman, Thinking, Fast and Slow, chapters 35–36. Product links: lesson 2 (The Psychology of Waiting) on the weight of how a wait ends; lesson 10 (honeymoon cliff) on budgeting for first impression; the cost track’s tiered routing and semantic cache are the engineering base under this lesson’s budget allocator.

✅ What this lesson wants to share