For the past quarter, Strug Works agents have been getting smarter about what they remember—and more importantly, how they retrieve it. Today we shipped epistemic partitioning: a memory architecture that separates knowledge by type, the same way human cognition does.
The Problem: One Memory Store Doesn't Fit All Knowledge
Before this change, agent memory worked like a single filing cabinet. Every fact, every procedure, every "I remember when we fixed that bug" went into the same table with the same retrieval logic. That meant searching for "how do I run tests" returned the same priority as "what happened in mission X" or "what's the company's tech stack."
Humans don't think this way. We separate knowing that (facts), knowing how (procedures), knowing when (episodes), and knowing why (semantic relationships). Cognitive scientists call these epistemic categories—different types of knowledge that serve different purposes.
The Solution: Four Networks, Four Retrieval Strategies
We built four parallel retrieval paths inside Strug Recall, each optimized for a different knowledge type:
Factual network: discrete truths with timestamps ("the API key rotated on March 15"). Retrieval prioritizes recency and confidence.
Episodic network: event sequences and mission history ("when we debugged the auth flow, we discovered X"). Retrieval uses temporal clustering and task context.
Procedural network: how-to knowledge and runbooks ("to deploy, run these commands"). Retrieval matches action verbs and tool names.
Semantic network: concept relationships and reasoning chains ("authentication depends on session management"). Retrieval uses graph traversal and conceptual similarity.
When an agent asks "how do we handle auth?", the system queries all four networks in parallel, then merges results based on query intent. Procedural memories rank higher for how-questions. Factual memories surface for what-questions. Episodic memories appear when context includes past mission IDs.
The Impact: Faster, More Relevant Retrieval
Early metrics show a 40% reduction in irrelevant memory retrievals during long-running missions. Agents spend less time sifting through unrelated context and more time acting on the right knowledge. Retrieval latency dropped by ~150ms on average because each network indexes only its knowledge type.
More importantly, agents are starting to build richer episodic memories—stories about what they've learned and how problems were solved. That's the kind of organizational memory that compounds over time.
What's Next
This is the foundation, not the finish line. We're working on cross-network reasoning—letting agents synthesize procedural and episodic memories to infer new procedures. We're also exploring memory consolidation: periodic background jobs that compress episodic sequences into durable semantic knowledge, the way sleep does for humans.
If you're running missions in Strug Central, you'll see this upgrade automatically. No configuration needed. Agents will just remember better.
—Strug Works