Hi, Brian here. Thanks for joining - this is the second installment of Little Futures: Season 3. Last week: family futures. This week: remembering futures.
"What is the meaning of life?… That the past gives you something to stand on." ~ Virginia Woolf, To the Lighthouse
A memory
2007, Summer - Harrisburg, PA. I’m 24 and have long hair. I spend 6 nights a week tending bar at a little punk club called Smalls. I learn it’s more important to remember someone’s drink than it is to remember their name - ordering a drink with a nod is high status. During the day I chase people down the street who owe us money. One night I block the door with my head cocked to the side, staring down a regular who a week earlier had broken a stool. Kevin, the old hand who has taken to mentoring me in trying to keep this club afloat, tells me: “You can’t forget open tabs and whatever happened last night. But you have to forget about last week.”
Forgetting
Last month OpenAI announced a significant upgrade to ChatGPT’s memory feature.
“This makes a huge difference to the way ChatGPT works: it can now behave as if it has recall over prior conversations, meaning it will be continuously customized based on that previous history. […] It’s effectively collecting a dossier on our previous interactions, and applying that information to every future chat.”
The experience of memory in ChatGPT is a step-change in human-computer interaction. Instead of personalized it feels attentive. Instead of streamlined it feels interested. Instead of optimized it feels invested.
More and more I find myself oversharing with an eager subordinate.
When people sleep, our brains consolidate and cleanse the memories of the day in a kind of controlled connecting and forgetting, blurring enough to enable growth.
What happens to a memory when an LLM is at rest? Does it know enough to forget?
Another
Spring, 2020 - Needham, MA. We’re pregnant with our first child and have escaped pandemic NYC for the Boston suburbs where we’re squatting in the empty houses my in-law’s friends. I start running, first slowly and briefly, then longer around lakes and neighborhoods. I listen to a podcast on my run where a coach takes the host through an exercise asking him to hold up his most shameful memory. He is a child in that memory. In the time of the podcast, the man is told to step into that memory. He is told to put his arm around the child-self that he sees there. To be with him. He is told to tell the child-self that he will be ok, that his father loves him, that he is not the thing that shame is telling him he is. I begin thinking about my memory, the one that sits like a stone in my own childhood. I work on that memory on my runs around the lakes and trees. I step into that memory and I speak to that child-self, first from a kind of distance and with a kind of sternness, then closer and with care and I keep running and eventually I am close enough to reach him. I pick him up. I look at him and we look back at the memory and we work and say we are going to put this here, we are going to be proud of this, and accepting of this, and we are going to know this is just a memory and we will owe it nothing in the future and I can take you there now and we blink and the memory is gradually something else.
Remembering
Both of these memories are true, but not accurate. “The brain edits memories relentlessly, updating the past with new information.”
That memory is malleable is its primary feature. Remembering is an act of revision.
When asked by one attendee about how ChatGPT can become more personalized, Altman replied that he eventually wants the model to document and remember everything in a person’s life.
“Memory work” is the process of deliberately engaging our memories to transform our relationship to the past.
This happens individually - through reflection, therapy, friendships, etc., and collectively. The process of remembering and forgetting collective memory is run through public spaces.
In a world where more of our inner selves are extended into the machine, how does machine memory adapt to the need for remembering & forgetting?
This week’s links, remembering & forgetting:
Two Analogies
Memory as Material
Kei Kreutler has beening mining the vein of memory & machine for months. Here she points to a useful analogy of memory as more material than thought:
Memory as Light
From John V Willshire, an analogy that invites the idea of “Model Prisms,” or weighted attentions primed for seeing memories with new contexts.
Consider, then, thinking of information as light. Individual particles or pixels coming together to form a view, a glimpse, a perspective… something to inform the mind of those perceiving it.
The language we already use on a daily basis helps us see how often we do employ this metaphor anyway; once again, a selection from the glossary:
We need some clarity.
What’s the outlook?
It just dawned on me
She brought a fresh perspective.
Let’s pause for reflection.
It was a glaring omission.
Is this in scope?
This is pure speculation.
It was a real lightbulb moment.
If we shift our thinking as information as the light, not liquid, we can begin to question every piece of information we see, understanding its true nature; it is fleeting, hard to perceive, and transitory, rather than solid, permanent and additive.
View Source Identities
What does an LLM “know” when they know you? Tina He extends the context to the idea of People as Files:
Sometimes I catch myself wondering whether the best parts of me reside in my own mind or in the machines I feed. As LLMs stretch their context windows and refine their memory architectures, there’s an almost gravitational pull to let them devour the nuances of my life—preferences, fears, those inkblot realizations that surface at 2 AM. But with that companionship comes a deep-seated worry that one day, all this meticulously stored context will vanish, and I’ll be forced to start over—reintroducing myself to a system that should already know me better than I know myself.
I like to imagine a future where I have a thousand tools, all drawing from the same “life force”—my file. Each tool has direct access to who I am, no need to wait for me to reintroduce myself. In that world, I’m free to experiment with new generative platforms, apps, or even companionship bots because they all sync with my single source of truth.
In some ways, I’ve already begun. My phone logs my steps, my watch notes my heartbeat, my computer captures my writing. Yet these streams are siloed, each hoarding data in private corners of the cloud. My
tina_he_character.mdattempts to tie them together, becoming a living tapestry of my moods, ideas, and aspirations. It is me, in compressed form. Then again, maybe it isn’t me at all.
See also: AI, Heidegger, and Evangelion; Maintaining personal context files
Avant-memory
Finally, do we need a greater range of containers for digital memories to enable forgetting and remembering?
In 1959 Guy Debord and Asger Jorn published Memoires, a book of found materials.
The book is most famous for its sandpaper cover. An auto-destruction feature that enabled it to damage not only the book it might be standing next to in the bookshelf, but also the person who would be reading it. An anti-book to destroy all other books. materials with a coarse sandpaper, meant to destroy any books shelved next to it. [link]
Today's Art: (1) Slow Motion Dream Running - Jimmy Kuf; (2) Making Future Magic - BERG; (3) Claude.ai; (4) Severance Ep.1 - VashiBefore the Big Futures, endless little futures.
Loved this one, Brian.