Ever had that problem that AI just forgets everything?

Ok, so I was working in ChatGPT on a long running project. Tweak after tweak. Step after step. Hour after hour. ChatGPT's screen is getting slower and slower. What the heck is going on? And — as a software nerd — how do I fix it?

Start a new chat — and now it's fast, but doesn't remember anything about the hours and hours we spent. Maybe it was the chat! Copy and paste everything from the old chat into the fast chat... slowwww again.

Try "Projects." Try uploading a document with the whole chat copy/pasted in. All the things everyone tries. So eventually I lost enough, and enough work was hidden away — essentially inaccessible — that I needed to figure this out. This tool is just too damn amazing for me to keep losing my investment in time and tokens to bloat. What the heck was going on?

Here's what's actually going on

So, here's what I found. In the same way people's brains don't hold every memory of everything for instant retrieval, neither can LLMs hold all of our past conversations in memory for instant retrieval. In fact, it's way worse than most people realize — a new session is a spotless mind with ZERO recognition of what came before. They seem magically smart, so well-trained, so jammed with every piece of esoteric knowledge, that they feel like they should know everything — including everything we discussed before. Well? LLMs know what THEY know, but that does not include learning anything about you or what you talked about. The models are trained, the weights are fixed, the model is released. The hours and hours you and I put into chatting with AI — they have zero effect on the model.

There are tricks, though. ChatGPT — and now Claude.ai — gathers a few pages worth of facts about you over time. The things it thinks are important. Your name, your job, a paragraph or two about what you recently worked on. When you "transfer memory" from one LLM to another, that summary is what gets transferred. Not a lot of depth. Not a lot of detail.

That few pages of notes, though — that's it. Start a new session, and behind the scenes those notes get sent to the server at the beginning of the conversation.

ChatGPT and Claude.ai also have a second trick (if you know to enable it): basic word search on recent chats. Frankly, that's a baby step in the right direction, and it's today's state of the art. It gives the LLM the appearance of having a clue. Otherwise? Nothing. Not a thing about your project, your decisions, or anything else you've ever talked about. Period.

The dream was simple

Wouldn't it be great though to just open a new chat and type continue and have the session just pick up where you left off? Give the AI a little nap. A little refresher. Get it back up to 100%?

That was my dream. I just wanted to pick up where I left off. Without the slowness. Without the dumbing down that happens when the chat gets too long. I wanted it to just remember what we were doing, what we tried, what we discussed, what our conclusions were. I wanted to stop answering the same questions over and over. I wanted to give my AI a real memory.

So I built it

Well? Experimented. A lot. I mean — a lot. The AI itself prodded me along in my quest, suggesting the things it needed to be able to "rehydrate" — (its phrase!) — back to full freshness (mine!).

In the end, the bottom line was: make all the past chats searchable in such a way that the LLM could intuitively reach out and pull in the answers. Organize past chats so that the LLM could trust it could reach into its own past and easily and quickly grab the things it already answered. And, even better, grab the things I already answered.

I finally got there. I've got my continue_session now. Once you experience what AI can be like when it can seamlessly pull from everything we've ever talked about, you never go back.

Then it got weird

And here's the craziest, possibly best part. All that history? Yeah, it's portable now. The infrastructure that lets one AI pull from all its past conversations? It also works seamlessly on other AIs too. That wasn't a goal — that was a hey I wonder if I can... Yes, I could.

All my ChatGPT history? Claude has it now. All my Claude history? ChatGPT has it. Turned out to be cool enough that every AI I use — ChatGPT, Perplexity, Mistral, Claude, even Claude Code — has access to all of it.

So now what

That's when I realized this shouldn't just be mine. I built it for myself, but once it worked, it became obvious other people would want it too. So we made ContinueVault into a product.

Free tier is real and useful — seven days of rolling history. Pro is $19/mo for full history plus the cross-platform stuff. I'd rather charge enough to keep building than give it away and watch it die. If you've hit the same wall I did, give it a try.