Building hierarchical multi-agent swarms with local persistent memory

Over the past 3 months we've been running an autonomous coding agent swarm across 140 GitHub repos. Here are the core architectural lessons:

1. State Isolation Matters

Shared memory contexts pollute model attention spans rapidly. Keep agent memories partitioned by role (Planner, Coder, Reviewer, Tester).

2. Guarding against token runaway

Always enforce exponential backoff and max cycle caps on recursive agent sub-loops.

export interface AgentMemoryContext {
  shortTermScratchpad: string[];
  episodicTaskHistory: TaskSnapshot[];
  immutableSystemPrompt: string;
}

Would love feedback on how others are handling SQLite vs pgvector for local agent scratchpads!