Meta just shipped a model that remembers your last 128 million tokens and built the chat interface to prove it matters.
The Summary
- Meta released Muse Spark, a new foundation model with 128 million token context and multi-session learning that builds persistent memory across conversations
- The meta.ai chat interface got rebuilt with artifact rendering, canvas editing, and deep citation features that actually show you where the model pulled information
- This isn't just a bigger context window. It's infrastructure for agents that learn your work patterns over months, not minutes.
The Signal
Context windows have been the AI industry's storage wars. 128k, then 200k, then a million. But Muse Spark's 128 million tokens isn't just bigger. It's different architecture. Multi-session learning means the model builds a persistent understanding across separate conversations. You don't re-explain your codebase every Monday morning. The model remembers.
The real story is what Meta built on top. Simon Willison tested the new meta.ai interface and found artifacts that render interactive outputs, a canvas mode for iterative editing, and citations that link directly to the training data sources the model referenced. Not vague attribution. Actual footnotes.
"This is infrastructure for agents that learn your work patterns over months, not minutes."
The citation feature matters more than it sounds. Current models hallucinate sources or give you generic "according to research" hedges. If Muse Spark can point to actual documents in its training set, that's verifiable output. That's the difference between an intern who makes things up and one who shows their work.
Key technical specs:
- 128 million token context (roughly 96 million words)
- Multi-session learning across conversation boundaries
- Deep citation to training sources
- Canvas editing interface for iterative refinement
The timing connects to Meta's broader agent play. They've been running AI agents on Instagram and Facebook for months, learning interaction patterns at scale. Muse Spark gives those agents memory that persists. An agent helping you draft social posts doesn't just remember this session. It remembers your brand voice from three months ago.
Hacker News reaction split between impressed (287 points, 301 comments) and skeptical. Fair. Meta's track record on sustained product focus is mixed. But the technical foundation is real. If the model can actually maintain coherent memory across 128 million tokens without degrading into mush, that changes what you can build.
The Implication
Watch what developers build with multi-session learning in the next 90 days. The first wave will be obvious: coding assistants that remember your entire project, research tools that track your reading over weeks. The second wave will be agents that don't just complete tasks but learn how you work and start suggesting what you need before you ask.
If you're building agents, this is your new baseline. Context isn't about fitting more into one prompt. It's about models that learn your patterns over time and get better at predicting what you need. The companies that figure out how to store and query long-term agent memory first will own the next layer of Web4 infrastructure.