Supervisor Memory

[antawiryawan]

Hi All,

How can I add memory to a Condition Agent? It is typically used as a supervisor. I am using Agentflows V2.

[KeepALifeUS]

For supervisor memory in Condition Agents, consider an external state approach:

Pattern: External State for Supervisors

// Supervisor state (outside agent node)
const supervisorState = {
    routing_history: [],
    agent_outputs: {},
    current_context: null
};

// Before each routing decision
supervisorState.current_context = {
    user_input: userMessage,
    previous_routes: supervisorState.routing_history,
    available_agents: ["agent1", "agent2", "agent3"]
};

// After routing
supervisorState.routing_history.push({
    decision: selectedAgent,
    reason: decisionReason,
    timestamp: Date.now()
});

Why This Works for Supervisors

1. Memory persists across routing decisions
2. Full visibility into why routes were chosen
3. Debuggable — inspect state to understand supervisor behavior
4. Flexible — add any context you need

Implementation in Flowise

1. Use a Function node before the Condition Agent to manage state
2. Pass relevant history as input to the Condition Agent prompt
3. Use another Function node after to record the decision

The key insight: supervisors need memory of their own decisions, not just conversation history.

More on state-based coordination: https://github.com/KeepALifeUS/autonomous-agents

[xXMrNidaXx]

Supervisor memory is a great feature request! Here is how we have approached this at RevolutionAI (https://revolutionai.io) when building multi-agent systems:

Shared memory patterns:

1. Central memory store — All agents read/write to a shared vector DB. Supervisor has read access to everything.
2. Hierarchical memory — Supervisor maintains high-level state, delegates detailed memory to worker agents.
3. Event-driven sync — Agents emit memory events, supervisor aggregates and maintains global context.

Implementation tip: Use a simple key-value structure for supervisor memory:

{
  "global_context": {...},
  "agent_states": {"agent1": {...}, "agent2": {...}},
  "task_history": [...]
}

The key is giving the supervisor just enough context to make routing decisions without overloading it with every agent detail.

What specific supervisor memory patterns are you trying to implement?