discussion: Agent memory (sharing in workflows) #740
Description
Maestro's current linear output -> input memory model is too rigid for complex workflows. It prevents dynamic collaboration and efficient state management between agents, especially in parallel tasks.
Proposal: Introduce a flexible memory store scoped to a single workflow run.
Key features should include:
-
Shared Context: A primary memory store accessible to all agents in the workflow for sharing state and intermediate findings.
-
Scoped Context: The ability to create private memory spaces for parallel branches to prevent race conditions, with results mergeable back into the main context.
Benefit:
This will enable more sophisticated agent patterns, such as multiple agents collaborating on a task in parallel by reading and contributing to a shared list of findings in real-time. This is fundamental for building advanced, collaborative agent systems efficiently.
Source: #459