DEV Community

Piotr Sarna
Piotr Sarna

Posted on

I built searchable memory for my ChatGPT, Claude & Gemini history

We have all had genuinely useful conversations with ChatGPT, Claude, or Gemini and then watched the good ones scroll into oblivion. I kept solving the same problem twice because I could not find that one thread from three weeks ago.

So I built Backscroll: export your history from each assistant, upload it once, and it becomes searchable by keyword and by meaning. You can also ask a question and get an answer pulled straight from your own past chats, cited back to the exact conversation.

How it works

  • Import your ChatGPT / Claude / Gemini export (the file each lets you download).
  • It chunks + embeds every message (Supabase pgvector, gte-small 384-dim, via a keyless edge function).
  • Search is hybrid: Postgres full-text keyword + vector similarity, merged.
  • "Ask your history" runs retrieval, then an LLM (via OpenRouter) answers with citations to the source chat.

Stack

Next.js + Tailwind, Supabase (Postgres + pgvector + Auth, RLS on every table), OpenRouter for the LLM. No background recording, no screen capture — just the exports you choose.

There is a no-signup demo so you can try the search on a sample archive first: https://backscroll.xyz/demo

Would love feedback from devs who live in AI chat tools — what do you go back and look for most? Code snippets, decisions, prompts?

Top comments (0)