From Hard Stops to Silent Drift: What I've Learned About AI's Memory Problem

Sep 18, 2026 · by Terry Peterson · claudeworkflowmemory

When I started using AI intensely with all my work about 18 months ago, I quickly ran into what I called AI’s short-term memory limit: The context window. If you have many brief, one-off chats with AI, you likely never encountered this. But stick with one convo long enough and most AI chats would hit a hard-stop. Dead end.

What you may not realize, the way most AI chats work is every time you send something to AI, on the backend it’s actually resending the ENTIRE convo back across the interwebs so whichever GPU is about to get fried with your next request knows what’s being discussed.

The first workaround was projects. On platforms offered projects, you can run multiple separate chat conversations within a project; the project just groups them together. Each chat can still individually run out of context. Add text files, PDFs, CSVs, etc to the project, and these project files acts like your AI’s long-term memory.

I started keeping a text file of critical things I wanted the AI to never forget. I found it helped with hallucinations, which are more prevalent the more a context window fills up, and bonus, I didn’t have to describe as much in each session. From time-to-time I would update the text file and replace the one sitting in the project with the new one.

Then some platforms started attacking the hard-stop problem directly. Manus was the first one I used that never seemed to dead-end. Instead, it auto-compacts the conversation as it goes. So you can literally keep one chat going indefinitely (or until your credit card dies, one of the two). But here’s my concern: if you gave a conversation hard facts, data, or rules early on, and it keeps rolling past a compaction point, you have no real visibility into how much of that original information survived intact versus how much was summarized, re-summarized, and re-summarized. Keep the conversation going further, and you have no idea how much was actually retained from the earlier parts of the convo.

What if you’re having a detailed business convo about a P&L statement. In the first prompt it has every P&L row and column to analyze. Thirty questions later, that P&L got summarized into “we make money… I think… what were we talking about?”. But AI being AI, it ‘wants’ to make you happy so it will keep answering your questions with seemingly solid answers. Cue… hallucinations.

This rolling context window is only subtly different, but it represents a monumentally scarier failure mode than the old hard stop. It’s not that “the AI forgot.” It’s “the AI silently reweighted and effectively redecided what it remembers and which parts were important.” For anything with hard constraints, e.g., legal specifics, financial figures, or exact rules; that’s a real risk, and it’s not always obvious when it’s happening.

If you use browser-based AI Chats, what most people think of when they think ChatGPT, Claude, Gemini, etc. you can still achieve the same control over long-term memory. Open up any text editor, you don’t need anything fancy. Type out what would actually hurt if the AI misremembered it. BTW…If you’re a Claude Code user, check out my other posts about managing context windows and building the /handoff skill.

A few ideas for your text file:

Hard numbers and facts. Anything you’d wince at seeing wrong — quotes, dates, exact figures. Example: “The budget is $177,000, not $200K that you’ve rounded up before. Always be precise with dollar amounts!”
Standing rules. Things you’d otherwise repeat every session. Example: “Never mention commission, my name, or any form of contact info when drafting property descriptions.”
Who you are, briefly. One or two lines of context so the AI stops re-guessing. Example: “I’m an experienced solo broker/owner working mostly with luxury sellers. Stop suggesting what I have my ‘team’ do, or to ‘ask my broker for advice’.”
Words that mean something specific to you. If a term you use has meaning to you that could be different elsewhere. Example: “When I say ‘the crew,’ they are my subcontractors, not an in-house team or staff.”
What NOT to do. The mistake you’ve already had to correct once. Example: “Keep all blog post drafts under 500 words because I want my videos to be 2min or less.”

Update it when something changes. Keep the text file saved somewhere convenient so it’s easy to upload the text file to a one-off convo, or add it to a project.

This is basically a manual version of where the AI chat industry is leaning. Claude’s own memory system, Anthropic Dream, and Grok Bot’s memory handling for instance, retain facts across conversations automatically… at their discretion. It’s a more automated cousin of this (what I think to be) more exact manual system. Greater effort for greater reliability.

But AI is smart, it can figure out what to remember, right? My high-schooler can also do more advanced math than I… but do I trust his memory on what to get from the grocery store? “Nuff said,” Stan Lee.