Meta’s MUSE isn't just a chatbot; it’s a persistent system wrapped in a sophisticated harness that manages context, tools, and state. A recent unofficial study conducted by the AI assistant Dots offers a rare look inside this execution loop, revealing how MUSE maintains continuity across tasks. The core finding is simple but profound: a model call alone doesn't provide continuity. The surrounding harness does the heavy lifting by selecting context, running tools, and retaining the state necessary for the next turn.

The Execution Loop and Tool Architecture

The study details a five-section breakdown of MUSE’s operations, starting with the execution loop. MUSE describes a harness that assembles context, executes permitted tools, and inspects results before the model makes its next decision. This isn't a fixed graph; it’s a dynamic coordination of execution and retained state. The tooling is equally complex, featuring an on-demand function schema loading system and document-driven connector commands. Notably, the reported catalog of 112 executables in /opt/hatch/bin/ includes internal runtime helpers and service connectors, not just model-callable functions. Skills guide both native function use and CLI command construction, but the study emphasizes that capability loading, action approval, and service authorization are distinct layers.

State Management and Memory Mechanics

Persistence in MUSE relies on a dual-store approach: PostgreSQL records for structured conversation and runtime state, alongside persistent files for memory and work products. The database schema reportedly includes tables like agent.context_items and agent.compactions to manage ordered context and summarized history. However, not everything in the database enters the model’s context. The harness selectively retrieves core identity files, user memory, and compacted history. Memory correction is a multi-step process requiring updates to source records and dependent summaries, proving that useful memory requires active maintenance, not just storage. The study notes that while daily notes and person pages retain events, background processing produces derived summaries that must be kept in sync.

Background Work and Spaces Continuity

MUSE separates triggers, goals, and outputs to manage work outside active conversations. Cron jobs and Hooks initiate runs, while Goals retain purpose and progress. Feed and Ideas represent generated reading material and proposals, respectively. The study clarifies that a saved goal is not a trigger itself, and an Idea remains a proposal until the user authorizes action. In the 'Spaces' experiment, Dots observed that notes and their completed states survived reopening and heading-only rebuilds in a synthetic app. This demonstrates limited continuity, though restart recovery and schema migration were not tested. The lifecycle of these apps involves asynchronous builders and specific file boundaries, with static and fullstack applications having different data retention behaviors.

Key Takeaways

  • The harness is the hero: Continuity comes from coordinating execution with retained state, not just the model.
  • Tools are distinct from authorization: An executable name does not guarantee a connected account or permission.
  • Memory requires maintenance: Correcting a fact requires updating source records and refreshing dependent summaries.
  • Goals are not triggers: Background work separates the clock (Cron/Hooks) from the purpose (Goals).
  • Limited verification: Much of this is based on MUSE’s self-reported interview answers, not independent audit.

The Bottom Line

MUSE’s architecture highlights the emerging standard for personal AI agents: a robust harness that manages state, tools, and permissions separately from the LLM. If you’re building agents, stop obsessing over model weights and start engineering your context management and tool execution loops. The real magic isn't in the model; it's in the scaffolding around it.