Memory & Prompting

Juno uses ephemeral session memory to retain context across interactions. This memory feeds into prompt construction for the AI assistant and improves interface responsiveness without relying on persistent storage.

Memory Model

Memory is scoped per session and tied to the active wallet. When a user connects or switches, memory is reset.

Stored fields:

{
  address: string,
  signals: Signal[],
  tokenRisks: Record<string, TokenRisk>,
  filter: string,
  promptHistory: string[]
}

  • signals holds recent signal objects (type, confidence, timestamp)

  • tokenRisks maps token addresses to known risk flags

  • filter tracks the current UI signal filter

  • promptHistory stores the last few assistant interactions

Memory is stored in client state, no backend persistence or database.

Prompt Construction

When the assistant is triggered, Juno assembles a prompt using current memory + interface state.

Example output:

This prompt is sent directly to the language model. No on-the-fly blockchain queries are made at prompt time. All data is preloaded and scoped.

Behavior

  • On wallet disconnect: memory is cleared

  • On signal update: memory is patched in place

  • On assistant reply: promptHistory is appended

  • On refresh: memory is rebuilt from scratch

PreviousToken Risk IndexNextJuno's Decision Model

Last updated