Why I Built loom-db
Why I Built loom-db
Most AI agents have the memory of a goldfish. They process tokens, generate responses, and forget everything the moment the context window slides forward. That's not how brains work, and it's not how useful agents should work either.
The Problem
Traditional databases give you storage, not memory. There's a difference:
- Storage is passive — data sits there until you query it
- Memory is active — it decays, gets reinforced, and surfaces context automatically
When an AI agent needs to remember a user's preferences, a previous conversation, or a learned behavior, it shouldn't need a complex retrieval pipeline. The memory should just... be there. Relevant. Timely. Alive.
The Biology
I got obsessed with the Ebbinghaus forgetting curve. It's a formula from 1885 that models how memories fade over time:
A(t) = A(t-1) × (DecayRate)^(Δt)
Simple, elegant, and surprisingly effective when applied to a graph database. Memories decay naturally unless reinforced — just like synapses in your brain.
The Design
loom-db has three types of nodes:
- Concepts — semantic knowledge ("Rust is a systems language")
- Episodes — temporal experiences ("User asked about deployment at 3pm")
- States — emotional context ("User was frustrated")
Nodes connect via weighted edges, and activation spreads through the graph like a ripple. Stimulate one node, and its neighbors get a boost too. This creates intuition — surfacing relevant context without explicit queries.
The Edge
The whole thing is pure Rust, no OS dependencies. That means it compiles to WebAssembly and runs in the browser. Your AI agent's memory lives client-side, no server round-trips, no privacy concerns.
"What if your agent's brain was just a WASM binary in the browser?"
That's loom-db.
What's Next
- Dream Protocol — consolidation cycles that reinforce stable memories and prune weak ones
- Instruction tuning for the ayvu-talian model using loom-db as context
- Multi-agent memory sharing — agents that learn from each other's experiences
Check it out on GitHub. Stars welcome, PRs even more.