Exact Recall

Exact Recall is a personal tool for conversations and moments I don't want to forget. I chat with Claude, which asks follow-ups one at a time, then stores a structured summary I can search later. Still in active development — core logging and three search modes work; retrieval quality isn't measured yet.

Important conversations and moments fade — I forget details, misremember what was said, and generic note apps need too much structure in the moment to be useful.

• Conversational logging

  Problem

  Capturing a memory in forms was too much friction in the moment.

  Fix

  Claude interviews one question at a time, then finalizes summary, fields, tags, people, location, dates, and a completeness score — one-shot summaries missed who/when/where too often.

  Result

  Events land as structured records without a blank form.

• Follow-up reminders

  Problem

  Incomplete events stayed incomplete and memories faded.

  Fix

  SMS via Twilio and web-push reminders when completeness is below 75%.

  Result

  Thin records got a prompt to fill in before they went stale.

• Three search modes

  Problem

  Keyword search fails when you remember the gist, not the words.

  Fix

  Live SQL keyword, pgvector semantic search, and Claude answers with citations.

  Result

  You can find an event by exact words, meaning, or a question.

• Event dashboard

  Problem

  Logged events had no place to scan, filter, or reopen.

  Fix

  /events with filters, a needs-attention strip, and a transcript detail view.

  Result

  Incomplete and complete events are visible in one list.

• File attachments

  Problem

  Memories often include photos and documents, not just text.

  Fix

  Images, HEIC/HEIF, PDFs, and docs stored in Cloudflare R2 and linked to events.

  Result

  Evidence sits with the event instead of in a separate folder.

• Google OAuth

  Problem

  Preferences and events needed to follow a signed-in user.

  Fix

  Clerk Google OAuth with preferences synced to Postgres via webhooks.

  Result

  Account state survives across sessions without a custom auth stack.

In active personal use; still in progress. Interview-then-finalize and three search modes (keyword, semantic, AI-with-citations) are built — search quality across modes is not instrumented yet; that's the main open gap.