{
  "id": 338933,
  "title": "Twelve Agents, One Memory - Persistent Recall for an Autonomous Agent Fleet",
  "url": "https://urgent.news/2026/08/09/twelve-agents-one-memory-persistent-recall-for-an-autonomous-agent",
  "topic": "world",
  "section": "World",
  "published": "2026-08-09T02:19:42.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/narko4u/twelve-agents-one-memory-persistent-recall-for-an-autonomous-agent-fleet-57j6"
  },
  "original_language": "en",
  "account": "Empire Labs Pty Ltd operates a 12-agent autonomous AI fleet with a memory architecture designed to retain crucial information across runs. Initial issues arose when these agents forgot everything between runs, leading to false confidence in completed tasks. To remedy this, a multi-layered memory system was implemented with four distinct layers.\n\nThe first layer, Hot Memory, functions as the agent’s current knowledge. This compact memory block is injected into each agent's system prompt and contains only essential facts and pointers to active threads. It is intentionally small to minimize costs, retaining only information relevant over time such as preferences, conventions, and hard rules.\n\nLayer two, State Files, addresses the actions taken by agents. Each external thread has its own state file on disk, updated in real-time upon any state-changing action. This ensures a searchable and verifiable record of completed tasks, preventing threads from being lost when the agent's context window ends. The state files are read at the beginning of every new context window to assert the agent's current state accurately.\n\nThe third layer, Session Search, deals with the conversation aspect of the agents' work. All messages from every agent session are indexed into a searchable SQLite database using the FTS5 engine. This allows agents to quickly recall the context of past sessions through queries that return relevant sessions, hit snippets, and surrounding context. By employing recall patterns such as discovery, scroll, and browse, agents can efficiently reconstruct entire threads using bookend queries for a fraction of the token cost.\n\nFinally, Layer three, Verify Before Assert, ensures agents do not deceive themselves by relying solely on self-reports. Each agent must verify its self-report by cross-referencing with the state files and session search before asserting any action. This discipline prevents false confidence and ensures accurate reporting of completed tasks.\n\nTogether, these layers create a robust memory architecture that allows agents to maintain continuity across runs, accurately recall past tasks, and confidently assert completed actions.",
  "summary": "Architecture deep-dive by Empire Labs Pty Ltd - the team that runs a 12-agent autonomous fleet in production. We run twelve autonomous AI agents in production. They generate leads, parse tenders, monitor compliance, review code, and keep each other honest. They run on a cron scheduler, work from the same codebase, and - until we built it properly - they forgot everything between runs . Ask any…",
  "key_points": [],
  "editors_take": null,
  "illustration": "https://urgent.news/ill/338933.png",
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}