Backpocket: An AI Agent for Everything You Save

What the survey showed

The goal was to bring a save back on its own, at the moment it's useful, without changing where I save things. First I wanted to know if this was just me, so I surveyed 19 students and working professionals.

1. Capture89%save things to come back to✓ Works
2. Revisit24%reopen more than 30% of saves✕ Breaks
3. Find18%found it quickly last time✕ Breaks
4. In time24%rarely or never too late✕ Breaks

Takeaway · Saving works. Everything after it doesn't, so a better place to save wouldn't fix this.

Where people save

People use about three places to save things, and two of them catch most of it.

• Screenshots88%
• WhatsApp chats65%
• Instagram Saved59%
• LinkedIn Saved53%
• Notes app18%
• Bookmarks12%
• Notion12%

Takeaway · Screenshots and WhatsApp lead, so the agent lives in WhatsApp and reads images as well as links.

Who it's for

Splitting the answers by who was responding changed who I built for.

• Rarely reopen what they save

  73%

  86%

• Save job posts

  82%

  43%

• Saved 6+ things last week

  64%

  29%

• Often lose a save to timing

  64%

  0%

Students and job seekersWorking professionals

Takeaway · Both groups forget. Job seekers lose more, because their saves have deadlines.

The job to be done: when an opportunity or a conversation is coming up, show me what I saved about it, and nudge me until I act.

What's already out there

Saving and searching with AI isn't new. But the tools I looked at either need a new app, or wait for you to come and look.

My read of how each product positions itself

Open space↑ It comes back to you↓ You go and searchRecallTapFoldThe SavedRicordaStashrSaved tabsBackpocket

← Lives in a new appLives where you already save →

See it in WhatsApp

You share to the agent like any contact. There are no commands to learn: you save, ask, or reply with what you did.

The 6 AM brief

Search only helps if you remember to look. So every morning, the agent reads my calendar and inbox and sends one message with what matters today.

Google Apps Script

Runs in my own Google account at 6 AM.

→↓

Calendar and Gmail

Today's events, plus sender, subject and one line from confirmation emails.

→↓

The agent matches

Finds saves that fit each meeting, and jobs each email confirms.

→↓

One WhatsApp message

Nothing useful today? No message at all.

What's in the brief

• Before today's meetings

  Saves that help with who I'm meeting, with the one point worth reading.

• Still waiting on you

  Saved job posts I haven't applied to, deadlines first, numbered so I can reply.

• Applied, from your inbox

  A "thanks for applying" email marks the matching job as done.

How it works

Every message runs through the same pipeline. Claude handles each judgment through its own tool with a strict schema, and every answer is validated before it's used.

WhatsApp Cloud API

Every message reaches a signed webhook.

Google Apps Script

Posts calendar events and confirmation emails at 6 AM, signed the same way.

↓

Next.js on Vercel

Checks the signature and the allowed numbers, answers Meta right away, then does the work in the background.

↓

The agent, on Claude

1. 01

   Route

   Is this a save, a question, a list request or an update?

   Claude

2. 02

   Read

   Pulls the page text, or the image from WhatsApp's media API.

   Cheerio, Media API

3. 03

   Write a card

   Title, summary, key points, topics, and company, role and deadline for jobs.

   Claude vision

4. 04

   Embed

   Turns the card into a vector so it can be found by meaning.

   OpenAI embeddings

↓

Supabase

Postgres with pgvector for cards and meaning search, private Storage for screenshots.

Reply on WhatsApp

Vector search finds the 5 closest saves, Claude answers from them, and the reply quotes your original message.

• Eight focused tools

  Routing, cards, answers, lists, screenshot checks, prep picks and email matching each get their own Claude tool.

• Checks its own output

  Every model response is validated. If it fails, the agent retries once, then replies with a friendly error.

• Safe by default

  Signed requests only, an allowlist of numbers, no repeat saves when Meta retries, and each person's data kept separate.

Key decisions

What broke

I used it on my own saves from day one. Every fix came from something going wrong.

• Questions saved as notes

  What broke

  "Can you share all the internships I haven't applied to" got saved as a note.

  What changed

  Claude reads what each message is for, not its punctuation.

• Answers too long

  What broke

  First answers restated the whole save in 500+ characters.

  What changed

  They lead with the answer, in about 200.

• Wrong message quoted

  What broke

  A slide sent seconds after a hidden job link got merged into it, so answers quoted the wrong message.

  What changed

  A screenshot only attaches if it actually shows that link.

• Two actions in one message

  What broke

  "1 skip, 3 applied" only recorded the skip, and an application was lost.

  What changed

  Each item gets its own action.

• LinkedIn posts cut short

  What broke

  Link previews stop after a few hundred characters.

  What changed

  It reads the full public post or job description.

• Reminders were slow

  What broke

  Reminders alone were slow to turn into applications.

  What changed

  They're numbered, so one short reply is enough.

Results

North star: the share of saves that come back at a useful moment and get acted on.

• Time to find a save

  10 to 30 min

  Under 1 min

• Job posts that came back

  ~5%

  100%

Before Backpocket With Backpocket

Before figures are my own estimates. Retrieval is for saved job posts; across all saves, 53% came back.

I've also acted on half the job posts I saved: applied, or skipped on purpose.

What I'd build next

• Next

  Test with 5 job seekers

  So far it's only me. The question is whether saving sticks past the first week.

• Then

  One-tap actions

  Applied and Skip buttons on every reminder, so acting takes a tap instead of a typed reply.

• Later

  Reels

  Read a reel's public caption, and keep the screenshot for everything else.