01 / The bottleneck
Vector recall answers what is similar. Agents also need what is related: who decided, with which tool, toward what outcome.
02 / The insight
Memory should be a graph — agents, memory, and knowledge as connected nodes — not a pile of embeddings.
03 / The build
A bench for relationship memory: entities and decisions linked across sessions, retrievable by relation as well as similarity.
04 / The launch
Bench experiment feeding Raven and the production agent setup.
05 / The outcome
Open research: relation-aware recall for multi-agent coordination.
06 / What I learned the hard way
The interaction itself becomes documentation when memory is traversable.