A second memory has two halves, and we want AI in both.
Recall is where we started: hybrid exact + semantic search over your notes, in any language, plus an MCP surface that lets AI assistants search and write memory with your permission. But today recall only ranks by similarity. We want it to understand intent — time, entities, and questions. Ask "the sneakers I bought last week — the details, and when do they arrive" and get the answer, not ten vaguely similar notes.
Capture is the bigger prize. Today capture means the user types a note. The future is zero-friction: with privacy as a hard constraint, Forgetwell tracks and structures your own digital history — orders, receipts, messages, browsing — so the sneaker question above has an answer even though you never wrote anything down. You forget freely; your second memory does not.
You will own the quality bar for both halves.
What you will do
- Own embedding and retrieval quality end to end: models, chunking, ranking, evaluation
- Make recall understand intent — time ranges, entities, and question-shaped queries — not just vector similarity
- Build zero-friction capture: turn messy real-life signals (orders, receipts, messages) into structured, queryable memory
- Treat privacy as a product constraint, not a checkbox: data minimization, user-controlled processing, on-device or local inference where it fits
- Build the evaluation harness that tells us capture coverage and recall quality got better, not just different
- Improve multilingual search and meaning-based recall on real user vocabularies
- Evolve the MCP tools so assistants can use memory safely and usefully
- Keep inference fast and affordable at single-binary scale
What we look for
- You have shipped retrieval, search, or embedding systems to production
- You have real opinions about privacy-preserving ML: on-device inference, data minimization, processing the user controls
- You can extract structure from messy real-world data — emails, receipts, notifications — and make it queryable
- You measure before you tune, and you distrust vibes-based quality claims
- You can work close to the product: latency budgets, UX, and privacy constraints
- Bonus: intent parsing and query understanding, multilingual NLP, on-device or small-model inference, MCP or agent tooling