What is Agent Memory?

Models are stateless by default: every call starts from nothing but what you hand it. Agent memory fixes that in two layers. Short-term — or working — memory is the running transcript of the current task: the plan, the tool results, the last few turns. Long-term memory is what survives the session: facts written to a store and retrieved when relevant, so an agent asked to file something like last time can look up what last time actually meant instead of guessing.

Velaris agents are multi-step LangGraph agents, so state is explicit: each step's output becomes part of the graph's state rather than being re-derived from scratch. That matters for the same reason context engineering does — memory is a budget, not a bucket. Long-term recall is retrieval, not accumulation: relevant facts get pulled in when a task needs them, the same way the agent queries the capability mesh for tools instead of carrying every schema in its prompt.