AI Memory vs. Custom Instructions vs. BootFile: Who Actually Knows You?

Every major AI platform now ships two personalization layers: a custom-instructions box you fill in yourself, and a memory system that accumulates facts about you automatically. In 2026 both got serious upgrades. But they capture different things, neither one travels with you, and there's a third layer — how you actually think — that both mostly miss.

Here's how the pieces fit together, platform by platform.

What memory looks like in 2026

ChatGPT rolled out a new memory system this month built on what OpenAI calls a "dreaming" architecture — instead of only storing individual saved memories, it synthesizes context from your past chats. You get an editable memory summary, and a new "memory sources" view shows which information personalized a given response — past chats, saved memories, custom instructions, files, or connected Gmail. The new system is live for Plus and Pro users in the US, with other plans and regions following.

Claude turned on memory for all users in March 2026. It automatically captures preferences, ongoing projects, and working-style signals, and acknowledges when stored context is influencing a response. Projects get their own isolated memory space, separate from global chat memory.

Gemini groups everything under Personal context: saved info (standing instructions), learning from past chats, and — for Google AI Pro and Ultra subscribers in the US — Personal Intelligence, which can pull from your Gmail and Photos to ground answers in your actual life. It's off by default. Gemini's memory has also drawn criticism for treating saved info as a high-priority instruction in every response, surfacing personal details in conversations where they don't belong — a relevance problem Google has said it's working on.

Memory collects facts. It doesn't capture reasoning.

Notice what every one of these systems is built to do: accumulate facts about you. Your job, your car, your project names, your stated preferences. That's genuinely useful — an AI that knows you drive a 2019 Outback gives better tire advice.

But watch what happens when you ask any of these tools a hard question. Memory doesn't tell the model whether you want the full landscape before a recommendation or the bottom line first. It doesn't know whether you process new ideas through analogies or through structured frameworks. It doesn't know whether pushback earns trust with you or burns it. Those aren't facts that surface in conversation — they're patterns in how you think, and memory systems aren't designed to extract them.

There's also a structural catch: every memory system is a silo. The profile ChatGPT builds about you doesn't exist in Claude. Claude's memory doesn't follow you to Gemini. The more you invest in one platform's memory, the higher the wall around you gets.

Custom instructions: the right layer, left blank

Custom instructions are the layer where reasoning style actually belongs — a standing description of how you want the AI to work with you, applied to every conversation. Every platform has one: ChatGPT's custom instructions on all plans, Claude's profile preferences, Gemini's saved info.

The problem isn't the feature. It's the blank box. Most people who open it write something like "I'm a product manager. Be concise." — true, but it describes your job and a formatting preference, not your cognition. The instructions that would actually change your results are the ones you'd never think to write down, because nobody examines their own reasoning style from the outside.

Where a BootFile fits

A BootFile is the content for that box — a structured, nine-section profile of how you reason, built from an eight-archetype cognitive assessment rather than from a blank page. We've written before about how it differs from custom instructions generally; the short version is that it covers the layer memory can't see and the blank box never captures: how you process information, what kind of answers earn your trust, how you want to be challenged, and what failure looks like.

The three layers compare like this:

	Memory	Custom instructions	BootFile
What it captures	Facts about you, accumulated passively	Whatever you think to write	Your reasoning style, extracted by a structured assessment
How it's created	Automatic, over many chats	Blank box, you improvise	Three-minute quiz → generated profile
Portable across platforms?	No — each platform is a silo	Only if you rewrite it per platform	Yes — formatted for each platform's instruction system
Can you inspect it?	Partially (summaries, sources)	Yes	Yes — it's a file you own

These layers stack rather than compete. Memory handles the facts. Your BootFile handles the cognition. Together, the AI knows both what's true about you and how to actually talk to you.

The bottom line

Memory in ChatGPT, Claude, and Gemini got meaningfully better in 2026, and it's worth turning on. But memory accumulates facts, not reasoning style — and it locks everything it learns inside one platform. Custom instructions are the layer built for reasoning style, and they're only as good as what you put in them. A BootFile fills that layer deliberately: one profile of how you think, portable to every AI you use.