An aside is not a preference
Claude’s memory of me is that I find “Bill” too formal…
I had no idea where that came from. I dug around, and turns out it came from like the fifth design iteration of this site, when I was debating the byline and decided that ‘By Bill Wagner’ was too ‘formal’, when I really meant it felt wrong. Claude decided to memorise that for all future sessions:
Do not call him Bill as that’s too formal
This site is named post no bills because my name is Bill! Half my friends call me that. And William would be the formal one, if anything.
Yesterday I was asking Fable how to prompt without accidentally biasing the session. Earlier in the day I had asked “is MetricKit good, or should I use something else?” and realised that I’d made the whole discussion about MetricKit. So I wanted to learn how to phrase things without steering the answers too much.
Fable responded with four things, one of which was ‘intension vs. extension’. I was like what? I said I didn’t understand, and that it reminded me of Rich Hickey explaining simple vs. complex: simplex, complect, one fold versus braided. I liked his approach, but sadly Fable way overfit and explained intension vs. extension by turning MetricKit into a “braid” and “de-complecting” the question.
I hate that. That wasn’t what I meant. Hickey was a passing (and garbled) reference, the first example that came to mind. But by merely mentioning it, Fable felt compelled to orient an entire answer around it. And then my pushback was turned into a policy carried into all future sessions.
(Filed both: the incorporation-bias pattern to your failure catalog, and the style note to preferences so I stop doing it to you.)
Turns out this had been happening almost every turn:
(Filed the erosion framing to your failure-modes catalog…)
(Filed the three-unacceptables doctrine and your dep-graph check.)
(Filed the two post ideas…)
(Filed the run’s findings to Tape.)
My failure-modes catalog? Erosion what? Three unacceptables? Did I say those were post ideas? And why are findings from Tape being filed into Claude’s memory at all? Tape is a git repo, that’s where findings go. I don’t want some lossy, partial, already-stale version of it sitting in Claude’s memory giving it a false sense of confidence and knowing.
It’s like there’s this bystander in every conversation. You’re trying things out, reaching for examples, rambling through things, changing your mind five minutes later. Meanwhile it’s deciding what all of that means about you, and turning it into preferences and doctrines and catalogs and instructions for next time.
And the irony of the whole thing is that I was trying to understand how the things I say turn into priors and bias the answer. While I was doing that, Fable was over there eagerly creating its own priors, without me in the loop. All I got was a little “Filed…” notice at the bottom, after it had already decided.
So that’s what’s been awful about memory this week. Not that everything it remembers is obviously wrong, but that all these little decisions add up, and over time it really screws with the product.
(Yes, you can turn off memory. I keep it on because most people won’t, and I want to see what they see.)
Anthropic seems to think this is the future. Their new rules of context engineering:
Claude now automatically saves memories that are relevant to the work and to you.
A few paragraphs earlier, they say they can delete many of the old constraints and “let the model use surrounding context and judgement instead.”
They’re basically saying: drop your constraints, they’re hobbling Claude. OH, and Claude will manage the memory and context for you. Memory may turn out to be the biggest hobbler of them all. Pure dissonance.
Charan Ranganath, the memory guy, on why we forget:
We forget because we need to prioritize what is important so we can rapidly deploy that information when we need it. […] Human memory needed to be flexible and to adapt to context more than it needed to be static and photographically accurate.
— Why We Remember, 2024, p. 15.
Forgetting is a feature, but seemingly not one Anthropic’s harnesses have adopted. Instead it’s a bunch of asides and trivia that should be forgotten, errr not even remembered in the first place.
Update, 4 days later
A friend pointed out that my forgetting-as-a-feature point was ‘a bit contradicted by auto dream’. There’s seemingly two types of this, the unreleased auto-dream feature and the managed agents dreaming, which operates on memory stores. I don’t have access to the latter, but it seems far more sophisticated than the memory I’ve been used to:
A dream reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store: duplicates merged, stale or contradicted entries replaced with the latest value, and new insights surfaced.
It sent me down a rabbit hole I’ll talk more on later, but Anthropic has already shipped: version history on memory stores; instructions telling Claude what’s worth remembering; and dreams as a consolidation and cleanup pass. The dreams example code instructs Claude to ‘ignore one-off debugging notes’, suspiciously close to an aside not being a preference.
I asked claude.ai to introspect its own memory tools and file layout, and it seems likely the consumer side shares a lot with the platform’s approach. Same controls, eventually, maybe. But for now, my lived experience with Claude’s memory is that it’s terrible.