When you remember something and when you imagine something, your brain may be using the same tools. I tested that idea on 6,854 real stories.
Most of us picture memory as a recording we play back. Research from the Schacter Memory Lab at Harvard suggests something different. Each time you remember an event, your brain rebuilds it from bits and pieces. And the brain uses the same rebuilding process to imagine things that haven't happened yet, like picturing tomorrow's job interview1.
This idea explains two things at once. It explains why we can imagine the future: we reuse pieces of the past. And it explains why memory makes mistakes. Because a memory is rebuilt rather than replayed, details can shift, gaps get filled in, and a story told many times can slowly change2.
So here is the question this project asks: if remembering and imagining use the same machinery, do real memories and made-up stories look the same when people write them down? Earlier research gives a hint about where they might differ. Real experiences tend to leave behind sensory details (what you saw, where you were, when it was). Made-up events tend to show signs of thinking, words like "I think" or "maybe"3.
A person writes about something real and memorable that happened to them in the last six months, plus a one-line summary.
A different person reads only that one-line summary and writes the story as if it had happened to them. Same event, but made up.
Months later, the original person writes about the same memory again, without looking at what they wrote the first time.
Below are real excerpts from the collection. Half are true memories. The other half were written by someone who never lived the event. Make your guess, then see what a simple computer model guessed. The model never reads the story itself; it only looks at counts like how many numbers, dates and "thinking" words it contains.
I measured every story in 25 ways. Some are simple counts: how many numbers, dates and times, names and "thinking" words (think, know, realize) it uses. Others use published lists in which thousands of people rated individual words. For example, those lists say how concrete a word is ("table" is concrete, "idea" is not) and how strongly it brings to mind each of the senses ("bright" is about sight, "loud" is about sound). Altogether those lists cover about 90,000 rated words.
The chart below compares real and imagined stories on all 25 measures. Blue dots on the right mean real memories have more of something. Orange dots on the left mean made-up stories have more.
Real memories lean on what people saw. Made-up stories lean on what people felt inside their bodies.
I placed every story on a map so that stories about similar things sit close together: weddings near weddings, hospital visits near hospital visits. Scroll down to explore it, and hover over any dot to see what that story is about.
Months after their first story, 1,310 people wrote about the same memory again. Comparing the two versions shows memory being rebuilt: same person, same event, different day.
This is the surprising part. You might expect the memories we go over most often to be the most stable, like a well-worn path. Instead, the memories people think about most, and care about most, change the most between tellings. That fits the idea that each time you recall a memory you rebuild it, and each rebuild is a chance for it to change. It also matches older research showing that vivid, emotional memories feel very accurate but aren't any more consistent than ordinary ones5.
Earlier research by the Schacter lab found that older adults include fewer specific details both when they remember the past and when they imagine the future. The two decline together6. That is strong evidence that both use the same machinery. The writers here are aged 18 to 55+, which is younger than in those studies, but we can still check whether memory and imagination move together as age increases.
After writing, each person answered questions about their story: how important the event was, how often they think about it, how stressful it was to write, and so on. This section checks how those answers connect to the way they wrote. That is 720 checks in total. Hover over any square to see the details.
These stories were collected online, not in a controlled lab study, so the results here are clues rather than proof. Each clue points to an experiment that fits one of the lab's main research areas.
Collect the same kind of real and made-up stories from people aged 18 to 85. Check whether the automatic word measures used here pick up the drop in detail that earlier studies found in older adults, and whether memory and imagination keep changing together in later life.
Ask people to recall an event more or fewer times, in more or less detail, before telling it again. Measure how much it changes, and whether imagined details sneak into the later version.
The split between "what it was like" writing and "what I thought about it" writing could become an automatic score for any story. Check it against brain scans taken while people imagine events, and against later mix-ups between what really happened and what was only imagined.
Lab studies show that a short exercise focusing people on specific details of past events boosts creative thinking afterwards7. Does it also make their imagined stories more "what it was like"? Here, personality (being open-minded) barely affected writing style, which suggests style can be shifted by training.
The stories. The Hippocorpus (Sap, Horvitz, Choi, Smith & Pennebaker, 2020, Microsoft Research): 6,854 stories, 1.8 million words, written by 3,640 online volunteers.
The word lists. Three published lists where many people rated individual words: how concrete each word is (Brysbaert et al., 2014; 37,058 words), how much each word relates to each sense (Lancaster Sensorimotor Norms, Lynott et al., 2020; 39,707 words), and how pleasant and intense each word is (Warriner et al., 2013; 13,905 words).
Nothing is invented. Every number on this page is calculated from these sources by a single Python script (analysis/build.py) that anyone can rerun.
The 25 measures, as defined in the script:
| Measure | What it means |
|---|
For researchers. Differences are reported as Cohen's d (paired dz for same-event pairs) with 95% confidence intervals. Links are Spearman correlations. "Reliable" means p < .01 (p < .001 for the 720-cell grid). The model is a logistic regression with 5-fold cross-validation grouped by writer (AUC 0.69 on the 25 measures, 0.78 on all words). Similarity between tellings is TF-IDF cosine similarity. The map uses TF-IDF, then SVD (80 dimensions), then t-SNE; topics come from k-means (k = 16).
Limits to keep in mind.