On Memory

Recollect — the recovery or recall of knowledge, bringing back to mind or memory

We live in an era where chatting with AI is an everyday thing.
But how do you remember your conversations with AI, the way you remember chats with a person?
Do you just open a new session and start from scratch every single time?

Lately, I've been juggling dozens of AI sessions across multiple projects simultaneously. Sometimes I deliberately break tasks down to open new sessions, or intentionally try to compact the conversation.

Because of this, I forget what I was even talking about after just 30 minutes. And then the needle-in-a-haystack search begins as I click through my old sessions one by one.

I don't think I'm the only one experiencing this. In fact, this is a recurring topic during our team's coffee chats—and I've noticed similar movements across the industry too.

Ever since OpenClaw came out, we've entered an era where everyday data overflows into context windows. "How much can we fit in?" is no longer the main bottleneck; instead, the memory problem of "how do we retrieve what's piled up?" has emerged.

Not too long ago, Andrej Karpathy also made waves with a single text file—llm-wiki.md. The idea was to map out a personal knowledge base in three layers: raw source, an LLM-organized wiki, and a schema, running it with three operations: ingest / query / lint.

Garry Tan's GBrain (github.com/garrytan/gbrain) seems to tackle the exact same void—"funneling meetings, emails, tweets, calendars, voice memos, and raw ideas into a searchable knowledge base, letting agents read and write before and after responding."

Some folks build up their context in Obsidian, Notion, or Linear and use AI to extract it.

These RAG methods are essentially semantic searches for resources. They just fetch the "resources" you need.

But I tried shifting my perspective a bit.

I often share thoughts on this topic with my teammates. We frequently gather around a single question: "What's been your biggest bottleneck using AI lately?" This particular day, the conversation started during a 1:1 coffee chat, and the story naturally drifted to this scenario.

"Where should we go for our coffee chat today?" "Oh, how about that place? The one with the orange sign on the main road, where we used to grab that fascinating orange coffee all the time."

Honestly, all we needed to remember was the shop's name, "Hammersmith Coffee." But these days, it feels like we increasingly construct sentences by stringing together remembered contextual references rather than using proper nouns. It almost seems like adapting to conversing with AI is changing how human memory recall works.

Lately, I've been obsessed with finding words that perfectly explain this phenomenon. In philosophy, they call it Salience. It's when specific attributes, scenes, or contexts stand out and act as a bridge for recall. The orange sign, the main road, the orange coffee—these are reference points pulled out instead of the actual "shop name" resource.

During that coffee chat, my teammate shared an interesting study.

It was the Nun Study, where David Snowdon tracked 678 nuns from the Sisters of Notre Dame for over 30 years. The linguistic markers found in short autobiographies the nuns wrote in their early 20s apparently had an incredibly strong correlation with Alzheimer's diagnoses 60 years later.

The marker was Idea Density—how tightly propositions (verbs, adjectives, adverbial phrases, prepositional phrases) were packed into 10 words.

• Low-density subgroup: About 80% developed Alzheimer's

• High-density subgroup: Only about 10% developed it

"I saw a red coat today" versus "I saw a red coat today, and it made me happy because it reminded me of the one I wore when I was five." The latter records a single scene by layering multiple references and emotions. What Snowdon observed was that writing with this kind of densely woven language throughout their lives was linked to building up cognitive reserve in the nuns' brains.

Of course, the brain's biological reserve and an external knowledge system architecture operate on completely different mechanisms. Still, the hint this study gave me was crystal clear: densely woven records support recall even if parts of them fade. Couldn't we apply this exact insight to our external AI session memories?

So, I figured the storage method that maximizes Salience is adding "Narrative" to the "Resource."

But "Narrative" is a bit too vague. Thinking back on that coffee chat, I broke down the dimensions that act as bridges for recall into six concrete axes.

• Coinage — Specific naming/neologisms for that scene. — "Hammersmith Coffee"

• Actor — Characters/agents involved. — "The teammate I had the coffee chat with"

• Temporal — Time references. — "Used to all the time"

• Emotional — Emotional markers. — "Fascinating to drink"

• Cognitive — Cognitive/insight markers. — "Oh, right"

• Singularity — One-off/unique events. — "Orange sign"

Instead of just storing the resource ("what was done"), I index these six axes alongside it. Think of it as overlaying a dense reference web onto your AI session memory.

Reading up to this point, you might be thinking, "Isn't this just metadata tagging?" Where my approach differs is this: while tags are classification labels like topic=AI, tool=Claude, these six axes are recorded as reference points that act as "bridges" during recall. The goal isn't categorization, but structuring it so that a single fuzzy clue can guide you back later.

You might also wonder, "Then why not just open a huge context window and stuff it all in?" That method is still powerful, but when dozens of sessions pile up over a month, how you signal to retrieve them becomes a much bigger bottleneck than how much you can fit. Each axis is basically designing a signal for retrieval.

The same goes for querying. Instead of using "exact keywords," you pull things up using Salience signals. The moment a fuzzy clue snaps into a past context and hits you with that "Ah, that!" realization—recall becomes recognition, not search.

There's an interesting asymmetry here. Standard RAG makes writing lightweight (chunking + embedding) and reading complex (multi-queries, reranking, deduplication). But the mechanics of memory seem to run in reverse—there's a dynamic where enriching the writing makes the reading simpler. If you put effort into writing across the six axes, a single fragment like "orange sign" is enough when it's time to retrieve.

Once I started logging and retrieving my AI sessions this way,

• The cost of recall dropped noticeably

• Handoffs for tasks from a month ago took just seconds

• The context-switching bottleneck was reduced, so even when running multiple projects in parallel, I no longer struggled with "finding that one chat from back then"

Looking back, the spark for crafting this kind of skill ultimately started at that coffee chat. Without my teammate's story about the nun study, the breakdown of "Resource + 6 Narrative Axes" would never have emerged.

"What's been your biggest bottleneck using AI lately?"—our DelightRoom Foundation team brings up this question on a regular basis. We constantly carve the discomforts we each find, and the rough hints that flow from them, into actual skills and protocols. We value the time spent observing each other's bottlenecks and talking them through just as much as the time spent building tech.

As we enter the great AI era, time spent having these kinds of conversations feels increasingly precious. If you're someone who enjoys this kind of wavelength, you might be curious about our team.

References

• Snowdon, D.A. et al. (1996). Linguistic ability in early life and cognitive function and Alzheimer’s disease in late life. JAMA. — PMC Summary

• Karpathy, A. llm-wiki gist — github.com/karpathy

• Tan, G. GBrain — github.com/garrytan/gbrain