article --summary
The summary
The piece argues that failures usually called memory problems come from three distinct context types: workflow state for resuming runs, durable memory for facts carried across sessions, and retrieved knowledge for the current request. A LangGraph checkpoint can resume a run but cannot decide what is worth remembering, so the application must enforce scope, validity, correction, and deletion for memories rather than leaving it to the prompt. Retrieval brings in outside evidence, but semantic similarity alone does not show that a source is current, authoritative, or permitted for the caller.
The author treats LangGraph and LangChain as building blocks rather than systems of record or final permission boundaries, so clear ownership shows where to investigate when something breaks. Readers are encouraged to design recovery paths alongside the normal flow, such as resuming after a failed tool without repeating side effects, superseding stale memories, excluding deleted documents before generation, and tracing which layer produced a wrong answer. The suggested starting point is one small workflow, with each of the three context paths tested separately before they are combined in the agent loop.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.