personal_asset

The Third Time Someone Says the Same Thing, Is It Nagging or Evidence?

Fixing the dedup logic in my own AI memory tool, I found the same similarity rule works in opposite directions for technical facts versus personal preferences — a mistake I also make with people all the time.

同一句话说了第三遍,是唠叨还是证据?

I built a small tool for myself on the side: whenever I talk something through with an AI and land on a judgment or a lesson learned, it gets logged automatically into a library I can browse later, so I don't have to organize notes by hand. Last week I was fixing the "dedup" logic in this tool — the part that stops the same thing from getting logged fifty times and clogging the library — and ran into a problem I hadn't anticipated.

Dedup is simple in principle: take a new entry, compare it against what's already in the library, merge it if it looks alike, keep it as new if it doesn't. I'd been running this logic for almost a month. That day I found the same ruler, applied to two different kinds of content, gave opposite answers.

The first kind is technical judgments — things like "skipping this one field in the request silently corrupts the data." If this kind of content shows up again and again, that's just noise. It should be merged, keep only the clearest version, and move on — extra copies just take up space and dilute what's actually useful.

The second kind is personal preferences and attitudes — a way of doing things I keep endorsing, a choice I keep saying I'd make again. When this kind of content repeats, it isn't noise at all — it's evidence. How many times it came up, how much time it spanned, whether it stayed consistent across different situations — all of that is telling you this is real, not a passing mood.

Running the dedup script and reading through the logs that day, I finally caught on: apply the technical-judgment ruler to the second kind of content, and the script sees the same attitude showing up again and treats it as "just repeated noise," merging it down into one line. All the evidence — how often, over how long — gets washed out. What's left in the library is one flat, dried-out conclusion, with no way to tell if it was a passing remark or something someone actually held onto.

The fix was to split the two paths: technical judgments still get merged when they look alike; anything about a person first gets its occurrences, sources, time span, and original wording gathered together, and only then does the system decide whether to update an existing judgment or log it as something new. The count and the span are themselves information worth keeping — they shouldn't get swallowed the moment "dedup" enters the picture.

Only after I'd fixed the code did it hit me — this is exactly how I react to people. When someone says the same thing two or three times, my first instinct is usually "yeah, I got it, you don't need to repeat it," and I file it away as nagging. But flip it around: something only worth saying once is usually something you don't care that much about. What someone truly cares about, what they actually want you to remember, is exactly what gets repeated over and over. I'd been measuring with the ruler built for technical facts, on something that needed a different ruler entirely.

Repetition in facts is noise — merge it. Repetition in a person is a signal — keep it. The line isn't drawn by what the content looks like; it's drawn by whether you're looking at a fact or at a person's attitude.

Once the dedup script was fixed, I didn't feel like I'd repaired some system. Instead, the first thing I thought of was: has anyone said the same thing to me a third time recently — and did I actually take it in, or did I just quietly filter it out as noise?

Sources