Memory isn't a keyword search problem. It's a connection problem.
This week we merged spreading activation retrieval into Strug Recall, the memory layer that powers every agent in Strug Works. The feature brings graph-based traversal to context retrieval, letting agents discover knowledge by following relationships rather than just matching words.
What Changed
Before this change, Strug Recall used BM25 ranking—a traditional information retrieval algorithm that scores documents based on term frequency and rarity. It works well for direct queries but struggles when the knowledge you need is one connection away from what you asked for.
Now we run a dual-trigger system. BM25 still handles the initial retrieval, surfacing documents that match the query terms. Then the Synapse algorithm takes over, spreading activation energy across the knowledge graph to surface related concepts, connected learnings, and contextual neighbors.
The algorithm works like this: initial matches receive activation energy. That energy spreads to connected nodes—memories linked by shared tags, temporal proximity, or explicit references. Nodes accumulate activation from multiple paths, and the top-scoring results combine both direct relevance and network centrality.
Why It Matters
Agents need context that matches how engineering teams actually work. When sc-backend investigates a database issue, it shouldn't just retrieve memories tagged 'supabase.' It should also surface the authentication fix from two weeks ago that touches the same tables, even if the original query didn't mention auth.
Spreading activation makes that possible. It turns memory from a lookup table into a web of knowledge, where retrieving one fact can illuminate its neighbors. That's closer to how humans recall information—and it's what autonomous agents need to build genuine understanding over time.
What's Next
This is phase three of our memory infrastructure work. Phase four will focus on decay and reinforcement—teaching Strug Recall which memories stay relevant and which should fade. We're also exploring temporal weighting to prioritize recent context without losing valuable long-term patterns.
The Synapse algorithm is live in production now, running on every memory read across the platform. We'll be monitoring retrieval quality and adjusting activation decay rates based on real-world agent behavior. If you're using Strug Works, your agents are already benefiting from smarter context retrieval.
Better memory means better decisions. We're just getting started.