How to Audit Your AI Memory in 10 Minutes

The monthly habit that keeps your AI working

You asked Claude for help with a project and the response felt... off. Not wrong exactly. Just tilted. Like it was answering a question you asked four months ago.

Maybe it referenced a tool you stopped using. Maybe it framed suggestions around a business model you abandoned in January. You shrugged it off. Rephrased. Got a better answer the second time and moved on.

But the bad answer wasn’t a glitch.

Claude was working off a stale profile of you the whole time. Old job context, dead projects, outdated positioning, tools you’ve swapped out. All of it sitting in a memory layer that loads into every conversation you have. Every single one. And it was bending your responses toward a version of you that doesn’t exist anymore.

The Uncle at Thanksgiving

Claude kept pitching me a $5,000 service offering. Every business task I brought to it. “Based on your Done-For-You AI Systems package...” followed by suggestions framed around pricing tiers and client onboarding for a service I’d brainstormed for 45 minutes on a Tuesday night back in November. Never built it. Never mentioned it to a single person who might actually pay for it.

Claude remembered it like I’d been running the thing for six months.

My uncle does this. He’s 82. Every Thanksgiving he walks up, puts his hand on my shoulder, and asks how the baseball card collection is going. I stopped collecting when I was 12. I’m 42. Those Upper Deck cards, which oddly enough are still in a box in my basement, but I couldn’t name a single card in there if my life depended on it. In his mind, I’m still the kid at the kitchen table sorting rookies from commons with a price guide.

He’s going to ask again next year.

Claude was doing the same thing, except Claude was using that bad information to shape every response it gave me. Content suggestions built around an offer I didn’t have. Strategy framed for a service model I’d abandoned. Audience targeting pointed at clients I wasn’t serving.

Eight Wrong Things

When I finally ran a full audit, here’s what I found rotting in there: a dead service offering still listed as active, an old X handle, a subscriber count that was off by 600, four killed projects Claude still thought I was building, a target audience description from six months ago, and my old newsletter name showing up everywhere.

Eight wrong things. All shaping every response I got. For weeks.

I cleaned it up in 10 minutes. The difference was immediate. Responses the next day were tighter, more relevant, built around my actual situation instead of a ghost from last fall.

There are two ways to do this. I’ll walk you through both. (this is a Claude walkthrough, but copy+paste this into your favorite AI and ask it to do the same)

Option 1: The 2-Minute Scan

Settings. Capabilities. “View and edit memory.” That’s the path.

Click it and you’ll see a summary of everything Claude has stored about you. Read it. Delete what’s wrong. Edit what’s shifted. Close it. Done.

Do this every couple weeks. It catches the obvious stuff, like a wrong job title or a tool you dropped two months ago. Fast and surgical.

The limitation: you’re reading a flat list. You’ll catch things that are clearly wrong but you’ll miss contradictions between entries, and you’ll miss deeper context that Claude has picked up from your conversation history that doesn’t even show in the summary view. That layer needs the second approach.

Option 2: The 10-Minute Conversational Audit

This is what I did.

Open a new chat and ask Claude to walk through everything it remembers about you. One category at a time. I use seven: work and role, active projects, tools, audience and business, personal context, preferences, then contradictions.

Claude surfaces what it has. You confirm or kill each item. Batch the corrections at the end and apply them all at once.

Why does this catch more? Two layers. Claude pulls from your explicit memory edits (stuff you’ve told it to remember) and from the accumulated context it has synthesized from months of conversation history. Two different things. The manual settings page only shows you one.

The Seven Categories

Start with work and role. What does Claude think you do? Has your title or company or focus changed since you last paid attention?

• Active projects rot the fastest. I had four dead ones still showing as active. Four. If you brainstormed something exciting for one conversation and never touched it again, Claude probably still thinks you’re building it. That Tuesday night idea you had about a SaaS dashboard? Still in there. Active. Waiting for you to ask about it.

• Tools. I dropped tools and added new ones without ever telling Claude. It kept suggesting workflows around software I hadn’t opened in weeks. Check what it thinks you’re using.

• Audience and business is where things get expensive. Your positioning shifts. Your offer changes. Claude doesn’t know unless you tell it, and the old version keeps informing every piece of content and strategy it gives you. I was getting advice for a target audience I’d moved away from six months prior.

• Personal context is smaller. Family, fitness, hobbies. But it shapes tone and recommendations in ways you won’t notice until something feels off.

• Preferences and style covers rules you’ve given Claude about how to write, how to format, what to avoid. Some of mine were outdated. Probably some of yours are too.

• Bontradictions. After everything else, look at the full set together. Entries that conflict with each other across categories. This is where the conversational approach earns its 10 minutes.

The Schedule

Manual check: every two weeks. Ninety seconds. Scan and cut.

Full conversational audit: once a month. Or whenever something big shifts. New direction, killed project, new tools. Any of those should trigger it.

I spent 10 minutes on mine. Found eight wrong things that had been warping my responses for weeks. The next day everything Claude gave me was sharper. Tighter. Built for the person I actually am right now.

Your AI remembers everything you’ve ever told it. It just doesn’t know when to forget.

Go check your memory. I’ll bet you’ve got some baseball cards in there.

Ryan

P.S. If this kind of stuff is useful to you, I go deeper for paid subscribers. Walkthroughs, systems, the actual workflows I'm building and breaking every week. The free newsletter gives you the what. The paid side gives you the how, with the messy details included.