Field note · Long-horizon agents · June 2026
Agent memory is execution state, not a vector database
Semantic recall is useful, but a long-running agent also needs to know which branch it is on, what has been validated, which attempts failed and where it can safely resume.
The category error
Vector stores retrieve semantically similar fragments. They do not automatically reconstruct causal execution state. Similarity can mix successful and failed traces, old and current decisions or evidence from incompatible branches.
Memory as a state tree
For long-horizon work, represent goals, subgoals, checkpoints, decisions and branches explicitly. The active path becomes the agent’s working state; completed branches can be compressed and failed branches isolated rather than blended back into recall.
Write, manage and read
Memory engineering needs three separate policies: what is worth writing, how memories are validated or consolidated, and when they should be retrieved. Add contradiction handling, confidence, provenance, expiry and learned or policy-driven forgetting.
Operational application
A coding or incident agent should checkpoint after deterministic tests, record unresolved risks and resume from a known boundary after failure. This is closer to workflow state management than to chat history.
Executive takeaway
Persistent memory increases capability and liability at the same time. Organisations need retention, privacy, correction and deletion controls before treating agent memory as institutional knowledge.