Remembee. Remember what you heard today

Remembee curates what's worth remembering from your Bee, transcripts, or chat exports into a daily quiz on the web or by voice on Alexa+, repeated days later, as research shows recall beats rereading.

Comment

Inspiration

I hear far more in a day than I keep, from the reason a launch slipped to the shop a friend recommended over lunch. Bee, Amazon's wearable AI, already records those conversations and summarises them, yet you reread a summary, and rereading is a weak way to remember. Being asked works better: students who read a text once and then took three recall tests remembered 61 % of it a week later, against 40 % for those who studied it four times, and repeats spread over days beat cramming. I wanted that quiz written for me from my own day, and played where I already talk to Alexa.

What it does

Remembee turns what you heard today into a two-minute quiz and brings each question back days later.

  • Bring it in. Drop a transcript or chat export on remembee.ai (Zoom, Teams, Google Meet, Otter, YouTube captions, WhatsApp, iMessage, Slack and more), or let the remembee CLI read your Bee day.
  • Curation. Remembee picks the few things still worth knowing in a month and writes one question for each, resting on a single quoted sentence as evidence. It leaves out what goes stale or never was knowledge, such as a price, a to-do or who brings the drinks.
  • Review. You check every question, with the answers hidden, before it is saved.
  • Play. On the website, or by voice on Alexa with the quiz screen on an Echo Show. You answer in your own words, and Remembee grades the answer, quotes the sentence it came from, offers a free tip after a miss and asks again after 1, 3, 7, 14 and 30 days.
  • Privacy. A dropped file is held only until its questions are written, then deleted. With remembee add --use claude (or OpenAI, or Bedrock in your own AWS account) the transcript goes only to your own AI and never reaches Remembee.

How we built it

I built Remembee alone and from scratch, starting on 2 Oct 2026, as a TypeScript monorepo with close to 500 commits.

  • One curation engine, two hosts. The pipeline runs on the Strands Agents SDK, in a Lambda with Claude Sonnet 4.6 on Amazon Bedrock or unchanged on the user's computer with their own AI. It is source-neutral, so Bee is one adapter next to the file readers.
  • Grading in three tiers. Rules decide the easy answers in milliseconds, then Claude Haiku 4.5, then Claude Sonnet 5.5 with a 3.5 s timeout and a fallback, which keeps a grading call inside a voice turn's budget.
  • An MCP server for Alexa+ at https://remembee.ai/mcp (MCP TypeScript SDK v2, Streamable HTTP, OAuth 2.1 through Amazon Cognito), which ships its own quiz screen as an MCP Apps widget. One Cognito account serves the website, the CLI and the Alexa account link.
  • One AWS CDK stack with 123 resources across 18 AWS services, deployed with one command: CloudFront with WAF, API Gateway, nine Lambda functions, one DynamoDB table, S3 for the short-lived uploads, and Titan embeddings to resolve "quiz me on the standup" to the right questions.
  • The Alexa device path runs through my open-source Alexa Skill MCP bridge, with its agent (Strands Agents, Amazon Nova 2 Lite) on Amazon Bedrock AgentCore Runtime.
  • An evaluation loop runs the real curator on 30 synthetic materials, scores it against gold facts and eleven trap classes, and has Claude Opus 5.5 judge every question. A prompt change is kept only when it improves training and holds on validation.

Challenges we ran into

  • No access to the Alexa+ MCP Toolkit. It is a private preview, so I built and open-sourced Alexa Skill MCP bridge instead: per-user account linking, a "One moment." on slow turns via progressive responses, and the server's MCP Apps widget rendered to an image in a sandboxed headless Chromium (84 ms p50 in a local test) and shown full screen through APL. Every feature is generic.
  • The name. Alexa kept matching the name to an existing skill called "Remember Me" when I said "Remembee", so the skill opens with "Alexa, open memory quiz" for now, while the product and its welcome stay Remembee.
  • Spoken answers. On the Echo Show the letter B arrived as "bee", C as "see", and dates as resolved values like "XXXX-XX-18". The grader now maps them back to an option and asks back when an answer fits two.
  • My first device test failed at the start in most ways it could, so I rebuilt the voice flow the same day: starting a quiz now takes one sentence or none, and every reply says what to say next.
  • Latency. The MCP Toolkit quickstart asks for 500 ms round trips, which answers that need a model don't reach yet, so the budget is built for the 3 s best practice.
  • What is worth remembering. Early versions of the curator happily asked about a price that changes next week. Moving that took a written yardstick ("what was learned that stays worth knowing") and the evaluation loop.

Accomplishments that we're proud of

  • Remembee is live at remembee.ai, the CLI is on npm, and the quiz plays by voice on an Echo Show with its screen, built by one person in three weeks.
  • Grading agrees with hand labels on 98.8 % of 82 test answers, and rules alone decide 63 % of them. Over MCP, a spoken answer is graded in about 1 s at the median on a warm stack with 20 users answering at once.
  • On 10 held-out synthetic materials the judge's score for the curator rose from 4.06 to 4.61 out of 5 and precision from 0.76 to 0.80, while the four trap classes I targeted fell to zero.
  • 2,858 passing tests, and a stack that deploys with one command.
  • Open-sourcing the Alexa Skill MCP Bridge has enabled all Alexa+ track entrants without toolkit access to run an account-bound, multi-turn MCP app with a screen on an Alexa device.

What we learned

  • I couldn't judge a prompt change by reading a few outputs, since changes that looked better broke something elsewhere. One change per run, measured on a fixed corpus, was the only way that held, at about 71 USD of Bedrock (list prices) for the whole series.
  • Voice needs its own design. A reply that doesn't say what to say next ends the conversation, and answers arrive the way Alexa heard them.
  • Privacy decided the architecture early: keeping transcripts out of the database led to one curation engine that runs in the cloud and on the user's computer alike.

What's next for Remembee. Remember what you heard today

  • The Alexa+ add-on through the MCP Toolkit once I have access. Alexa+ then calls the MCP server directly and can show the same widget.
  • Reminders on Alexa when questions are due, so spaced repetition doesn't depend on remembering to open the app.
  • An installer for the CLI, and direct connections to Zoom, Slack and Google so nobody has to export a file.
  • Grading under 500 ms, with more rules and a smaller first model.
  • Releasing iOS and Android apps for also taking quizzes on phones, tablets and TVs
  • More seamless transcript uploads and question generation to reduce and eventually eliminate all manual effort from users.
  • Monetizing Premium features

Built With

Share this project:

Updates

Submission history