One of the hardest problems in building AI systems that feel genuinely intelligent is memory. Not the kind that lets an agent recall what you said five minutes ago in the same conversation—that's just context window management. We're talking about the kind of memory that lets an agent remember what you told it last week, or last month, and use that information to make better decisions today.
This week we shipped the foundation for that capability in Sabine, our AI partnership platform. Two new pieces went live: the store_memory skill and the Slow Path worker.
What Changed
The store_memory skill is exactly what it sounds like: a tool that agents can call to persist information they learn during interactions. When Sabine has a conversation with a user and learns something important—a preference, a piece of context about their work, a pattern in how they communicate—it can now write that down for future reference.
The Slow Path worker handles the async processing side. Memory operations don't need to happen in real-time—they can be queued, validated, and stored in the background without blocking the main interaction flow. This architecture keeps conversations fast while still building up a rich memory layer over time.
Why It Matters
Every AI product eventually hits this wall: the context window runs out, or the session ends, and the agent forgets everything. You start over every time. That's fine for a chatbot answering FAQ questions, but it doesn't work for a system that's supposed to be a genuine partner in your work.
Sabine is built on Strug Works, our autonomous product and engineering platform. When agents can remember what they've learned, they get better at their jobs. They make fewer mistakes, ask fewer redundant questions, and start to feel like they actually understand your context instead of just processing your prompts.
This is infrastructure work—not flashy, not immediately visible to users—but it's the kind of foundation that makes everything else possible.
What's Next
Having the ability to store memories is step one. Step two is teaching agents when to store them, what to store, and how to retrieve the right memories at the right time. We're working on retrieval strategies that go beyond simple keyword matching—context-aware memory search that understands intent and relevance.
We're also building out Strug Recall, the memory browser interface in Strug Central. Right now memory is purely agent-facing. Soon you'll be able to see what the system has learned about you, correct it when it's wrong, and explicitly tell it things you want it to remember.
Memory is one of those capabilities that unlocks a whole category of use cases we couldn't support before. Expect to see more of this work surface in Sabine over the next few weeks.