Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, and the headline feature for anyone building knowledge tooling isn't the coding scores. It's Active Memory — described by Meta as the ability to carry what matters across conversations and bring back details relevant to the task at hand. Paired with it is Active Focus, persistence on long-horizon goals: trying alternate approaches when blocked, adapting to new information, asking for clarification, and checking in before consequential actions. Meta claims state-of-the-art results on MemoryBench, FocusBench, SWE-Marathon, and DeepSWE. Muse Wonder is available through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output; Muse Wonder Full is gated behind Muse Code's $50 and $200 monthly tiers. Alexandr Wang told Axios the models are "particularly state of the art on personal and professional agentic work."
Why this is a boundary question, not a feature
Most teams shipping agents today own memory explicitly: an extraction step that writes facts or summaries, a store, a retrieval policy, a reranker, and some eviction or decay rule. You can inspect every row. You can delete a user's data and prove it. You can tune what gets recalled when precision matters more than recall.
A model-side memory makes that machinery partly redundant and partly invisible. The retrieval decision moves inside the weights and the provider's serving stack. You get better defaults for free — less re-explaining, fewer hand-tuned recency heuristics — and you lose the seam where you'd normally enforce tenancy, audit a recall, or explain to a customer why the agent surfaced a detail from six weeks ago. For consumer assistants that's a clear win. For a regulated internal knowledge base, an opaque recall path is a compliance problem before it's an architecture problem.
The pricing shape matters too. Memory you manage is mostly storage cost plus a small embedding bill. Memory the model manages is token cost on every turn, at $4/$20 per million.
The other direction of travel
Worth noting the contrast with what frameworks are doing. Microsoft's Agent Framework has been pushing the opposite way — its Python 1.18.0 release added shared vector-store abstractions, portable filters, an in-memory store, and Redis HASH/JSON and Azure AI Search implementations, with native contracts in `agent-framework-core` that deliberately avoid depending on Pydantic, NumPy, pandas, or Semantic Kernel. Microsoft labels these experimental and warns of breaking changes.
So one camp is standardizing an explicit, swappable, inspectable memory substrate. The other is absorbing memory into the frontier model. If you're architecting now, assume both exist in your stack within a year and keep the seam: treat provider memory as a cache or a convenience layer, not as your system of record. Keep your own durable store, keep writes idempotent, and keep the ability to answer "why did it remember that" without filing a support ticket.