I kept losing good answers in my AI chat history, so I built search for it
I use ChatGPT and Claude all day for work, and the thing that kept getting me was losing answers. I'd figure out something fiddly, move on, and a few weeks later I'd know the answer was buried somewhere in my history — but the built-in search is basically keyword matching, so good luck if you don't remember the exact wording you used.
So I built Backscroll. You export your ChatGPT/Claude history, drop it in, and you get real semantic search across everything you've ever asked. You search by what you meant, not the exact words. It's pgvector + embeddings doing the work behind a deliberately plain UI.
Honest status: it's live, it works, and it has zero paying users so far. I'm alright at building and pretty bad at distribution, which is most of why I'm posting here.
Two things I'd genuinely like your take on:
- Is "search my own AI history" a real pain for you, or do you not really care once a chat scrolls off the sidebar?
- For anyone who's launched a small dev tool — what actually got you your first handful of users? Reddit, HN, writing, DMs?
There's a demo so you can poke at it without signing up: https://backscroll.xyz
Happy to hear why it won't work. I'd rather get that now than in a month.
honestly this is a real pain, context gets lost way too fast
It really is. The annoying part is that the answer is usually still sitting there, so it's not lost, just unfindable, which somehow feels worse.
This is a real pain for me. The problem is not just finding an old answer, but recovering the context around it: what I was building, why a decision was made, what constraints existed at the time, and what the next step was supposed to be.
For small dev tools, I’d probably start with very specific use cases instead of broad “search your AI history.” For example: founders managing multiple projects, developers debugging with AI, or people who use AI as a long-term project memory.
The demo-first approach is good. I’d be curious if users search more for exact old answers, or for “what did I decide about X?” type context.
The distinction between finding the answer and recovering the context around it is the thing I keep underestimating. Constraints and the reason a decision was made are exactly what falls out, and a keyword hit on the answer doesn't bring them back.
To your question: I don't have clean data yet, but the searches I run myself are far more often "what did I decide about X" than "find that exact snippet". If that holds for other people it changes what the product should return.
Founders juggling several projects is the narrow case I'd start with.
That makes sense. If the most common search is closer to “what did I decide about X?” then maybe the result should feel less like search results and more like a recovered decision/context card.
Something like: decision, reasoning, constraints at the time, related conversations, and possible next step. For founders juggling multiple projects, that would probably be more valuable than just surfacing the old message.
The narrow “multi-project founder memory” use case feels strong because the pain gets worse over time, not better.
@danionescu You're describing the exact gap I keep hitting. Finding the message is easy. Recovering why that path made sense at the time is the hard part. The decision card format you sketched - reasoning, constraints, next step - feels closer to what actually gets lost. Still early days on whether it matters enough for people to change their workflow.
I hit this same wall and ended up building a custom memory layer for my agent rig (semantic indexing, nightly consolidation) because native search is a black hole. The manual export is your biggest adoption tax. The people with the most data are usually the least likely to spend 5 mins exporting it. Ever considered a browser extension to handle ingestion in the background? Making it zero-touch would be the killer feature.
The export being an adoption tax is the most useful thing anyone has said about Backscroll. You're right that the people with the most history are exactly the ones least likely to spend five minutes on a manual export.
A browser extension for background ingestion has been on the maybe list mostly because of the permissions conversation it starts. But "zero-touch" being the actual feature rather than the search is a reframe worth taking seriously.
Curious how your nightly consolidation ended up working, if you ever wrote it up.
Yes, real pain, but I’d validate it around trigger moments rather than “search history” generally. For first users, I’d look for people who already hit AI workflow breakpoints: “I solved this in Claude last month and lost it,” “I keep rebuilding prompts,” “our team repeats the same AI research.” Those are better than generic AI productivity spaces. I’m doing similar hand-to-hand work with a tiny AI usage utility, and exact-pain threads beat launch posts by a lot.
Trigger moments rather than the category, yes. "I solved this in Claude last month and lost it" is a phrase someone actually types. "Search your AI history" is not.
That also gives me something to search for instead of a community to post in, which is a better shape of work.
Good luck with the usage utility, and if exact-pain threads keep beating launch posts for you I'd take that as strong evidence.
Can't give you any advice on question 2 because I'm honestly stuck at the exact same stage with my own project, but l definitely wanted to answer your first question.
Searching through Al history is a massive pain for me too. It drives me absolutely crazy when you know you figured something out a few weeks ago, but it's completely lost and buried in the history sidebar because the built-in keyword search just sucks. Having to waste time searching for it all over again is incredibly annoying.
I haven't tested the demo yet, but this is 100% a problem that affects me daily. I'Il definitely check out backscroll.xyz and will make sure to drop you some more detailed feedback once l've played around with it. Cool idea, definitely keep going!
Answering question one is genuinely useful, so thanks. "You know you figured it out and it's buried" is the exact feeling I built this for, and hearing it described back is worth more than a signup.
Same stage as you on the second question, so no advice to trade, only sympathy.
If you do poke at the demo I'd rather hear what made you close the tab than what you liked.
This is a real pain, but I’d be careful with how you test it.
The risk is not that people say “yes, I lose useful AI answers too.” A lot of people will agree with that.
The harder question is whether the pain is urgent enough that someone changes behavior, exports history, uploads it somewhere, and keeps coming back.
I would not solve the first-user path loosely in the thread because the right channel depends on which trigger you decide to test first.
If you’re open to it, share your email and I’ll put the tighter first-user path together properly.
Agreed on the risk. "Yes I lose answers too" is close to free to say, and I'd have counted it as validation.
The behaviour test you're describing is the harder one: export, upload, come back. That's three separate moments where someone can quietly decide it isn't worth it, and only the third one means anything.
I'd rather work it out in the open than by email, but thank you for offering.