Under the Hood: Migrating Genie's Memory to Supabase

At Genie, our mission is to build an AI that doesn't just chat, but collaborates. A true collaborator remembers. They remember the API endpoint you discussed last week, the architectural decision you made yesterday, and the bug you've been fighting all morning.

For a long time, Genie's memory was functional but limited. We relied on simple key-value stores and token-heavy context stuffing. It worked for short sessions, but as conversations grew, so did the latency—and the "amnesia."

Today, we're sharing how we completely re-architected Genie's memory system using Supabase and pgvector, achieving semantic understanding and sub-second retrieval times.

The Problem with "Simple" Memory#

Our initial implementation was based on what we call "Session Context." Every time you sent a message, we'd bundle up the last $N$ messages and send them to the LLM.

This had two major flaws:

1. Context Window Limits: We had to aggressively prune old messages. If you referenced a decision from 100 messages ago, Genie hallucinated or simply didn't know.
2. Cost & Latency: Sending massive context windows is expensive and slow.

We needed a system that could store infinite history but only retrieve the relevant bits. We needed Retrieval-Augmented Generation (RAG).

Enter Supabase & pgvector#

We chose Supabase for three main reasons:

1. Postgres is King: It's robust, reliable, and we already know how to query it.
2. pgvector: The ability to store vector embeddings right next to our relational data (users, conversations) is a game-changer. No syncing data between a primary DB and a separate vector DB (like Pinecone or Weaviate).
3. Real-time Capabilities: We can push updates to the UI instantly when a memory is formed.

The New Architecture: "Memory Bank"#

We introduced a new concept called the Memory Bank. It's not just a log of chats; it's a crystallized knowledge graph.

Here is the simplified schema we deployed:

create extension if not exists vector;

create table memories (
  id uuid primary key default gen_random_uuid(),
  user_id uuid references auth.users(id),
  content text not null,
  embedding vector(1536),
  importance integer default 3,
  created_at timestamptz default now()
);

When you talk to Genie now, a background process (our "Cortex") runs asynchronously:

1. It analyzes your message.
2. It generates an embedding using OpenAI's text-embedding-3-small.
3. It performs a cosine similarity search against your memories table in Supabase.
4. It injects the top 5 most relevant memories into the system prompt.

The Results#

The shift has been dramatic.

• Latency: Retrieval takes <100ms. Code generation starts almost instantly.
• Accuracy: Genie can now recall file paths, variable names, and project constraints discussed weeks ago.
• Developer Experience: Dealing with standard SQL to debug memory issues is infinitely better than black-box vector stores.

What's Next? 2026 Roadmap#

This migration sets the stage for our ambitious 2026 roadmap. We are moving beyond "Chat" and into "Agency."

Later this year, we'll be rolling out Active Memory—where Genie proactively updates its own knowledge base based on your code commits, not just your chat messages.

Stay tuned. The memory upgrade is live for all users today.