Website:
Job details:
Every AI you use forgets you the moment you close the tab.
I'm building the one that doesn't, and I'm looking for the person who builds it with me.
Emma is a personal AI with memory. You tell her something once, and she still knows it six months later, so you stop re-explaining your job, your relationship, your situation every time you open a chat window. The thesis is simple: the first AI that gets better the longer you use it.
The hard part, and the reason this is defensible, is that her memory is encrypted. Conversations are stored as ciphertext under a key derived from the user's own passphrase. We cannot read them. Most memory products are a feature bolted onto a chat app. Here the memory is the product: structured, visible, editable, and it fades unless you use it.
Where this actually is, honestly
React, TypeScript, Supabase. Postgres with pgvector, row-level security, edge functions. Client-side AES-GCM encryption, 768-dimension embeddings, cosine retrieval.
The memory system is built, and the chat loop is live. Five screens still render from sample data because the job that would fill them doesn't exist yet. 28 accounts, no marketing, no launch.
Early enough that your decisions shape the product. Not so early that there's nothing to work with.
The four problems waiting for you
The overnight pass that turns raw memory into patterns. Clustering, decay, consolidation. It's the highest-leverage unbuilt thing here, and it's what makes five sample screens real.
Retrieval that stays sharp as memory grows. Ranking on recency, strength, and salience rather than similarity alone. This is the genuinely research-shaped problem, and nobody has fully solved it.
The model layer. Currently mocked behind a flag. Open-weight and bring-your-own-key models are a first-class path, not an afterthought.
Mobile. PWA first, React Native when retention earns it.
Who I'm looking for
Strong TypeScript. Real comfort with Postgres, ideally with vector search. You've built something end-to-end and watched it break in production.
Applied experience over credentials. I care whether you've built retrieval that stayed good at scale, not whether you can derive attention from scratch.
Encryption and key handling is a serious plus, because it's the claim the whole product rests on.
And you push back. You'd be the only engineer, and I move fast and decide fast. Someone who agrees with everything is just an expensive pair of hands.
What I'm offering
Meaningful equity, four-year vesting, one-year cliff. Pre-seed, so this is equity-weighted, and I'll give you real numbers in the first conversation rather than make you ask twice.
Architectural ownership of a product where the architecture is the differentiation. No committee.
How this starts
No interview loop. A couple of weeks building something real together, part-time. You see the codebase and how I decide. I see how you think. We'll both know more than any interview would tell us.
If this is your kind of problem, apply.
Tell me about a system you built where the hard part was memory, retrieval, or trust.
Click on Apply to know more.