Lab
The errors a long-term memory can carry forward
Short answer
In a long-term memory, a mistake written once gets read again every time someone relies on it, and being written down gives it an authority it has not earned. The only serious defense is to make it visible: provenance, dates, statuses, corrections that replace without erasing, and spot checks.

A mistake in an ordinary chat disappears with the chat. In a memory that lasts, it stays. It gets reread in the next session, quoted in a summary, built into a decision, and each time it looks a little more solid, because it has already been used and nobody objected.
That is the downside of a long memory: whatever goes in by mistake lives as long as everything else. This article describes the main ways an error travels through time, walks through an example that spans six months, and lists what helps you catch it. It does not promise a memory without mistakes. It tries to show how to make them visible.
A mistake written once is read a hundred times
A conversation you forget costs nothing. A note you keep, if it is wrong, costs you on every reading. The mechanism is ordinary: the error is copied into the next piece of work, which becomes a source in turn.
Two forces work together. The first is the authority of the written word: what is on record seems to have been weighed, when the note may have taken thirty seconds. The second is repetition: the more often you read something, the more familiar it gets, and familiar things pass for true. We looked at one form of this bias in why found does not mean true. Here the question is what happens when the error has months ahead of it.
Five ways a mistake spreads
The transcription slip
A figure copied wrongly, a date misheard, a name misspelled. At the start it is a small slip. But once somebody uses it in a calculation, the error moves up a floor, then another.
The hypothesis that hardened into a fact
Someone writes “the client should sign off by Friday”. Three weeks later the “should” is gone and the note says “the client signs off Friday”. Nobody lied: the status got lost in a copy and paste or a summary.
The summary of a summary
A meeting gets condensed, then the condensed version gets condensed again for a monthly report. With each pass, nuances drop out and shortcuts settle in. By the third level you are reading a crisp claim that nobody ever actually made.
The outdated decision that keeps getting cited
A decision was replaced, but the old version is still the easiest to find: it is more detailed, or better titled. People keep quoting it. That is the case covered in what happens to the old decision when a decision changes.
The bias confirmed by its own traces
Someone believes a certain acquisition channel performs badly. They log the failures, forget to log the successes, then reread the memory and find a neat confirmation. The memory did not lie. It reflected what was put into it.
An example: a mistake that crosses six months
Imagine Nora, who organizes events as a freelancer. The example is made up.
March. On the phone, a caterer says: “For a hundred and twenty people it is possible, I will confirm next week.” In her call notebook, Nora writes: “Caterer: possible for 120, to be confirmed.” When she copies it into the project file, she writes: “Caterer: 120 guests OK.” The “to be confirmed” has fallen off.
April. A project summary picks up the line: “Caterer confirmed for 120.”
May. A new summary is written from the previous one, without rereading the original notes: “Caterer confirmed, deposit to arrange.” The deposit comes from a discussion that went nowhere.
July. Nora prepares a quote for her client. She relies on the summary, and the quote mentions a confirmed caterer.
September. The caterer writes that they are not available on that date. Nobody understands: everything was “confirmed”.
What lets the team trace it back is one small detail: each line keeps a reference to its source. The quote points to the May summary, which points to April’s, which points to the March note, which points to the call notebook, where the “to be confirmed” is still there. Without those references, the team would have hunted for someone to blame instead of finding a mechanism.
Why a long memory amplifies the problem
A short document gets reread in full, and the mistake gets seen. A long memory does not get reread: you search it for what you need and lean on whatever comes up. The error is only read by the people who use it, which means the people with the least reason to question it.
There is also a volume effect. The more notes there are, the likelier an error is to be copied elsewhere, and the costlier it becomes to find every copy. An error duplicated ten times is far harder to fix than a single one. And time works against human recall: in six months, nobody will remember the March phone call.
The habits that actually help
None is enough alone. Together, they make an error easier to see.
- Provenance. Every piece of information should be able to say where it comes from: said, written, imported, inferred. That is what makes the trace-back in Nora’s story possible. Depending on the information available, Nawame can keep when something was recorded, where it came from, and whether it has been changed or replaced since.
- Dates. A note without a date gives no way to tell whether it has been overtaken. It is also one of the signals described in how to tell that a memory has become obsolete.
- Statuses. “Said”, “assumed”, “decided”, “verified”: a status stops a hypothesis from hardening into a fact.
- Visible corrections. Correcting by replacing, without erasing the old version, leaves a trace of both the error and the fix. You can see what changed and look for what depended on it.
- Spot checks. Every so often, reread a few items picked at random and trace them back to their source. It is slow, and it works.
- No summaries of summaries. When you summarize, start from the sources, not from the last summary.
- Marked uncertainty. A “to be confirmed” written out costs three words and can save months of error.
If you want to measure how often these errors happen rather than guess, we describe our approach in how we measure memory errors.
Current limits
No memory is immune. The only serious defense is to make errors visible, and even that has limits.
Provenance only helps if it was recorded. Information copied by hand into a document with no source loses its trail, and no system can rebuild it. Statuses depend on the people who set them: a “verified” added too quickly is one more error that looks like a guarantee. Marking uncertainty everywhere has its own flaw: when everything is “to be confirmed”, nothing really is.
Spot checks take time, and people put them off. And nothing replaces the doubt of a person who asks “where does this come from?” at the right moment. A memory can make that question easier, but it cannot ask it for you. The Lab keeps studying how errors spread, and does not present the problem as settled.
Something to try: reread ten memories
Take ten items from your working memory or your reference document, at random rather than among the most recent. For each one, ask three questions:
- Where does it come from? Can I find the source, or only a summary?
- Does it have a status: said, assumed, decided, verified?
- Is it still valid today?
Write down the ones you cannot answer for. Those are your fragile spots. If one of them came from an import, read it twice: importing memory without importing hallucinations gives a method for that. And do the exercise again in a month, with different items.


