Your AI agent’s “memory problem” may be something else
An agent may need additional context to continue when a user returns the next day. A workflow may resume from The post Your AI agent’s “memory problem” may be something else appeared first on The New Stack .
Long-running AI agents require three types of context to function correctly: workflow state, durable information, and external knowledge. However, many issues are incorrectly labeled as "memory problems" and stem from various causes.
Workflow state holds where the agent stopped, nodes, tool results, approvals, and pending tasks. Agent memory stores facts, preferences, decisions, procedures, and validity for later use. External knowledge includes documents, chunks, embeddings, versions, and access scope.
Each context type needs a clear owner, lifecycle, and access boundary. While LangGraph and LangChain offer useful abstractions, developers must define these responsibilities explicitly. Keeping these responsibilities separate helps diagnose failures and operate long-running agents more smoothly.
Checkpoints let a workflow resume from the right point, but cannot define memory. Workflows can pause for approval, recover post-restart, or resume after tool failures. Checkpoints persist graph state under a thread identifier, including messages, tool results, and execution progress.
Retrieval provides current external evidence for the request. It uses keyword search, vector search, SQL, or a hybrid approach. Vector search ranks records by similarity but doesn't guarantee a record's currentness, authority, completeness, or permission for the caller. Retrieval should not replace agent memory, as support manuals, workflow decisions, and user preferences all have different owners, access scopes, and update rules.
Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.