Peak-End Rule: Users Only Remember the Peak and the End
Four session plots show how average score and memory score diverge; the product version of the cold-water experiment: spend the flagship model on first impression, put validation up front to protect the last step, and save freely in the middle
THE QUESTION THIS PAGE ANSWERS
ANSWER FIRSTWhat is the key idea behind “Peak-End Rule: Users Only Remember the Peak and the End”?
Four session plots show how average score and memory score diverge; the product version of the cold-water experiment: spend the flagship model on first impression, put validation up front to protect the last step, and save freely in the middle
Turn taste into a behavior the product can repeat. The useful outcome is not a nice opinion. It is a visible rule, a small example, and a way to tell when the experience falls below the bar.
Capture one before-and-after example that shows the quality bar without extra explanation.
Polish that improves the surface while leaving the user's uncertainty untouched.
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.
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.
One hand in 14°C water for a full 60 seconds, then done — towel off.
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.
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.
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.
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.
Turn the feeling in “Play the plots first · Four fates of the same session” into a judgment
“One 10-turn AI session, each turn scored 0 to 10.” points out that AI has lowered the bar for making something usable. The skill readers need is noticing what is wrong and turning that feeling into an actionable requirement.
Watch the user's next action, not just the surface
Turn “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…” into observable questions: does the user know what happened, what to do next, and how to recover from an empty or failed state? Does the hierarchy make the important information visible first?
- Memory score ≠ average score : users archive only the peak and end frames — nobody remembers the average
- Make the peak : spend the flagship model on onboarding and high-value moments; a cache-speed reply counts as a peak too
- Protect the end : validation up front, intermediate artifacts persisted — never make the last step the most failure-prone action in the whole chain
Pretty is not the same as usable
Apply “Budget done — now flow.” to a second screen or flow. Record one moment of hesitation and the user action after the change; observable behavior is stronger evidence than polish alone.
From “Play the plots first · Four fates of the same session” to “Unpack the source · That 1993 basin of cold water”
“Play the plots first · Four fates of the same session” grounds the problem in “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 an…”. “Unpack the source · That 1993 basin of cold water” then moves it toward “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…”. Together, they show that the lesson is not just a conclusion to remember, but a claim with conditions.
Carry the judgment into the next situation
For experience work, turn abstract impressions into user actions: did the person understand the state, find the next step, recover from an error, and want to continue?
- “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 an…
- “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…
- “The closing point”: Save freely in the middle : cheap models plus a semantic-cache floor — move the savings to both ends to buy memory score
The final “The closing point” brings the discussion to “Save freely in the middle : cheap models plus a semantic-cache floor — move the savings to both ends to buy memory score”. The useful thing to carry forward is knowing which judgments must be revisited when input, scale, or risk changes.
✅ What this lesson wants to share
- Memory score ≠ average score: users archive only the peak and end frames — nobody remembers the average
- Make the peak: spend the flagship model on onboarding and high-value moments; a cache-speed reply counts as a peak too
- Protect the end: validation up front, intermediate artifacts persisted — never make the last step the most failure-prone action in the whole chain
- Save freely in the middle: cheap models plus a semantic-cache floor — move the savings to both ends to buy memory score
I turned one judgment from this article into a small experiment I could run today. Knowing what to observe next is more useful than simply remembering the conclusion.
After reading this, I first looked for the conditions behind the idea instead of copying the method into a project. That order made the later trade-offs much clearer.
When this judgment reaches real work, which constraint should be added first? I am curious which step matters most between reading and the first practical attempt.
No discussion on this article yet.