Building Alma: a cognitive architecture for AI that actually learns who you are

I've been building Alma for months. It's a persistent memory and identity engine for AI — web app, MCP server, VSCode extension, REST API. I use it daily and I'm at the point where going back to vanilla Claude feels broken.

Here's what Alma actually does.

When you chat, a background processor fires every few messages. It extracts structured facts from the conversation — not a raw transcript dump, but discrete memories. Each one gets a category (preference, decision, project, fact), an importance score (0 to 1), and a confidence level (1.0 if you stated it directly, 0.7 if the AI inferred it). Duplicates merge automatically. Important things rise, irrelevant things fade.

That's the first layer. The second is episodes — compressed conversation summaries. Not "you talked for 45 minutes" but "debugged auth middleware, root cause was missing await, decided to add retry logic." Next time auth comes up, the AI already has that context.

The third is procedures — behavioral patterns the AI observes over time. "Checks error handling first in code reviews." "Prefers minimal comments in code." You never write these. They emerge from how you work.

All three layers feed into a 12-block identity system I call the Soul Engine. It separates who the AI is (stable identity) from how it communicates (style that evolves) from what it knows about you (context that auto-updates). The single highest-impact piece: the anti-patterns block, where you define what the AI must never do. Five rules there improved my output more than pages of positive instructions.

The tech: Cloudflare Workers, D1, Vectorize, R2, Durable Objects, Workers AI for embeddings. Hono framework. React 19 frontend. 2,622 tests across 149 files, 69 database migrations. Claude-only (Haiku free, Sonnet + Opus paid). Hybrid search combining keyword matching with vector similarity.

The honest part: I'm a solo developer. I launched on Product Hunt and TAAFT two days ago. I have a handful of paying subscribers and I'm figuring out distribution. The product works — the challenge is getting it in front of the right people.

If you use AI daily for work, the value kicks in around week two when the accumulated context starts compounding. The first session feels normal. The twentieth feels like a different product.

alma.olivares.ai — free tier, no credit card. Feedback on the product or the approach is genuinely welcome.