Recall rewrites the thing it retrieves. Every read is a write. If you take one thing from this post, take that, because it inverts the intuition most of us run on, which is that memory is a hard drive with an unreliable index. The index is fine. The file is what's unreliable, and it's unreliable in a specific, predictable direction.

The rats

In 2000, Karim Nader ran an experiment that should have been impossible under the old model. He trained rats to fear a tone by pairing it with a shock. Standard conditioning. Once the fear memory had consolidated, which takes hours and was thought to be a one-way process, the memory was supposed to be stable. You couldn't undo it with the protein synthesis inhibitors that block the formation of new memories, because it was already formed.

Nader played the tone again, without the shock, to bring the memory back up. Then he injected the inhibitor. The rats forgot. Not weakened, not extinguished with a residual trace that could be reinstated. Gone, as though it had never consolidated.

The interpretation is called reconsolidation, and it has replicated across species and memory types for two decades now. Retrieving a memory returns it to a fragile state, and it has to be re-stored afterward. Re-storage is a construction step. It uses the current context, the current mood, whatever else is active in the network. Then it writes that version back over the old one.

There's no separate original. There is only the most recent write.

The verb changes the speed

The human version of this has been known since the seventies, mostly through Elizabeth Loftus. In one study people watched a film of a car accident and were asked how fast the cars were going when they "hit" each other, or when they "smashed" into each other. The smashed group estimated higher speeds. A week later, asked whether there had been broken glass (there was none), the smashed group was more than twice as likely to remember glass.

A single verb, in a question, after the event, edited the event.

Loftus later planted entire false memories, the well-known one being "lost in a mall as a child," in about a quarter of adult subjects by having a relative vouch for it. Some subjects then elaborated the memory with details the experimenters never supplied. The memory grew.

None of these people were lying. That's the point. Subjectively, a reconstructed memory and a veridical one are the same experience. There's no flag on the file that says "modified."

Why compression

Here's the part I find clarifying rather than just alarming.

What's stored isn't the event. It's a sparse encoding: a few high-salience details, an emotional valence, a rough spatial layout, some links to other memories that were active at the time. On recall, a generative process takes that sketch and fills in everything else from priors. What restaurants usually look like. What your friend usually says. What a Tuesday is usually like.

This is lossy compression with a decoder that hallucinates the missing bits from a learned distribution. It is, structurally, a lot like a modern image codec, and it has the same failure mode: the decoder is confident. It doesn't output "unknown" for the parts it doesn't have. It outputs the most probable thing, at full resolution.

And then the reconsolidation step saves the decoded output back over the sketch. If you have ever re-saved a JPEG twenty times you know what happens. Each generation looks fine on its own. Compare generation one to generation twenty and it's a different image.

The trade

Why would evolution ship this? Because storage was never the bottleneck. Retrieval was.

A memory system that stored raw sensory experience would be enormous, and worse, useless, because you can't search it. What you need from memory is not a replay of the past but a fast answer to "what's this situation like, and what happened last time." That's a content-addressed lookup on gist. Gist is what you store because gist is what you query on.

The fact that it also lets you update a memory with new information is a feature. If you learn later that the person at the party was your cousin's roommate, not your cousin, the system quietly patches the scene. Most of the time that patch is correct. The system is optimized for the average case, and the average case is that the world is consistent and your priors are good.

The cost shows up in the tails: eyewitness testimony, arguments with your partner about who said what in 2017, and the specific, awful confidence of a memory that has been rehearsed a hundred times.

Rehearsal makes it worse

This is the counterintuitive one. Neisser and Harsch asked students the morning after the Challenger explosion to write down where they were and how they heard. Two and a half years later they asked again. The later accounts were vivid, detailed, and confidently held, and a quarter of them disagreed with the original on every major detail. Confidence was uncorrelated with accuracy. Shown their own handwriting from day one, several subjects said the original must be wrong.

The memories that get rehearsed the most are the emotional ones, so they get re-encoded the most, so they drift the most, and each re-encoding raises the felt certainty. The memories you are surest of are the ones that have been through the most write cycles.

There's a practical corollary for anyone who keeps a journal or a lab notebook, which is that the notebook isn't a backup of your memory. It's a different, better system, with a property yours lacks: the entries are immutable. You can't reconsolidate a page. When the two disagree, the page wins, and the interesting question is never "which is right" but "what did I gain by changing it." The answer is usually a better story, and occasionally an alibi.

What to do with this

Not much, honestly. You cannot opt out of the codec. But a few things follow.

Write things down close to the event, and when you later disagree with the note, trust the note. The note is generation one.

Treat vivid, confident, frequently retold memories as the least reliable, not the most. Especially your own.

And when you and someone you love remember the same evening differently, consider that you are probably both wrong in the same direction, toward whatever made a better story, and that there's no original left to check. There's only the two most recent writes.