Context engineering manages the information available to a model at each decision. Instructions are one input alongside task state, tool definitions, retrieved material and prior observations. Anthropic's engineering account describes selective loading and compaction as ways to manage this changing working set. context-engineering
Persistent memory retains information for future interactions. Its defining concern is a managed lifecycle of retained knowledge, not a particular vector database or the current prompt length. LangChain distinguishes thread-scoped state from cross-thread memory and discusses profiles, collections, and synchronous or background updates. langgraph-memory
Retrieval brings external evidence into a task; grounding ties a claim or action to that evidence. They are related but distinct: finding a relevant document does not establish that it supports the answer. The original RAG paper demonstrates combining a learned generator with retrieved non-parametric information in specific NLP tasks. Modern retrieval designs need not reproduce that architecture. rag
Reusable procedures package deliberately authored instructions and supporting resources for repeated tasks. They solve a context distribution and maintenance problem: keeping specialized guidance available without placing all of it in every prompt. This page is emerging because the cross-runtime packaging and activation contract remains less established than ordinary instruction reuse.