Memory stores prior conversation or state so later steps can refer back to it.
DefinitionWhat it means
Memory in an orchestration framework is the storage and retrieval of prior conversation turns, intermediate results, or user facts so that later steps in a workflow can reference them. It ranges from a simple rolling buffer of recent messages to a long-term store, often vector-backed, that persists facts across sessions and is retrieved when relevant.
Why it mattersWhy you should care
Without memory, every agent turn starts from a blank slate, which breaks multi-turn tasks and makes personalization impossible. Choosing the right memory strategy, short-term buffer versus long-term retrieval versus structured fact store, directly affects both the coherence of an agent's behavior and its token cost, since naively replaying full history gets expensive fast.
At a glanceSee it
Memory splits into a short-term context window and a durable long-term store, which itself divides into episodic, semantic, and procedural knowledge.
The write path gates each turn through a salience check before embedding it, while the read path retrieves the top matches and injects them — and can still surface a stale or wrong recall.
Where you see itIn the wild
- Chat applications that keep a rolling window of recent turns in context.
- Long-term memory stores that persist user preferences across sessions.
- Agent frameworks distinguishing short-term scratchpad memory from long-term retrieval.