915 B
2026-02-25 Session Notes
Predictive Memory Scorer Vision Discussion
Nicholai discussed the conceptual architecture of Signet's core mechanism: a memory prediction algorithm that learns to surface the most relevant memories based on context, sentiment, and moment. The conversation framed Signet not as a static memory system but as an inference engine that predicts optimal recall strategies.
Key proposal: implement a Rust-based transformer model trained locally and unique per user. This model would train progressively on accumulated memories and session histories, making real-time memory injection predictions at two critical moments: session start and on user prompt submit events. The inference pipeline pattern from MicroGPT was referenced as a potential design guide.
The vision positions this as Signet's north star feature—the difference between a tool that remembers and a mind that persists.