The gap between a tool and a teammate is not raw intelligence. It is memory.
An agent that forgets everything has to start from zero every time. An agent that remembers without structure becomes noisy and untrustworthy. Rush splits the problem into tiers.
The three tiers
1. User memory
This is global memory shared across agents. It holds durable facts about you: preferences, recurring constraints, and information that should not need to be repeated.
2. Agent memory
This is scoped to one agent. A research agent keeps its own domain-specific context. An email agent keeps its own. Knowledge stays useful without leaking everywhere.
3. Session memory
This is the current conversation. It is intentionally more disposable. The goal is to keep the active task coherent without promoting every temporary detail into permanent memory.
Why the tiers matter
A flat list of remembered facts works for one assistant. It breaks when you have many specialists.
Rush uses tiers so the system can answer three different questions cleanly:
- what belongs to the person?
- what belongs to this agent's ongoing work?
- what belongs only to the current moment?
Why provenance matters too
Useful memory is not just about storing a fact. It is also about knowing where it came from.
Rush records memory with source metadata so the system can tell whether a fact came from you directly or from an agent inference. That makes correction, trust, and deletion possible.
Memory is only helpful when it stays legible.