Shared vocabulary
Context Engineering Glossary
Context work gets easier when teams use precise names for the moving parts: what is selected, what is retrieved, what is compressed, what is remembered, and what is evaluated.
Key facts
Context Engineering Glossary
- Context engineering is the umbrella discipline; context curation is the selection step inside it.
- Retrieval, compression, memory tiering, structuring, and evaluation are distinct parts of the context pipeline.
- Definitions on this page are model-neutral and intended for team docs, onboarding, and evaluation rubrics.
- Stable vocabulary makes failures easier to assign to the right layer instead of labeling everything prompt quality.
"systematic optimization of information payloads"
Context engineering
The systematic design of the information payload an LLM receives during inference, including retrieval, processing, memory, tool results, instructions, and output constraints.
Context curation
The selection step inside context engineering: deciding which facts, examples, history, and rules deserve space in the active context window.
Context window
The bounded input and output token budget the model can attend to for a single generation.
Working set
The subset of available information assembled for the current task. A good working set is smaller than the corpus and more structured than raw history.
Retrieval augmented generation
A pattern that retrieves relevant external information and includes it in the prompt so generation can be grounded in explicit sources.
Long context
A model capability that accepts large prompts. It increases capacity but does not guarantee reliable use of every token.
Context rot
Performance degradation or inconsistency as input length grows, often caused by distractors, buried evidence, repeated instructions, stale history, or sheer length.
Prompt compression
A method for reducing prompt length while preserving task-relevant meaning, evidence, and constraints.
Chunking
Dividing documents into retrievable units. Good chunking follows semantic boundaries rather than token counts alone.
Reranking
Reordering retrieved candidates by a richer relevance model or task-specific criteria before context assembly.
Source card
A labeled evidence unit that includes source title, type, date, authority, retrieval reason, and the relevant passage.
Faithfulness
The degree to which an answer is supported by the supplied context rather than unsupported inference or prior model knowledge.
Answer relevance
The degree to which an answer addresses the user task at the right specificity and format.
Context relevance
The degree to which retrieved or assembled context contains focused evidence needed for the task.
Memory tiering
Keeping information in different stores, such as active prompt, short-term memory, summaries, vector indexes, and durable user preferences.
Output contract
The required shape of the model response: fields, citations, uncertainty language, validation steps, and prohibited claims.
How these terms fit together
Context engineering is the umbrella. Context curation selects the working set. Retrieval finds candidates from outside the prompt. Compression reduces selected material. Structuring makes roles and priorities visible. Evaluation checks whether the assembled context improves the task. Memory tiering keeps durable information available without dumping every past token into every prompt.
Sources Used
Research survey - 2025
A Survey of Context Engineering for Large Language Models
Frames context engineering as retrieval and generation, context processing, and context management, with a large taxonomy of implementations.
Research paper - 2020
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Introduces the RAG formulation that combines parametric model knowledge with non-parametric retrieved memory.
Research paper - 2023/2024
Lost in the Middle: How Language Models Use Long Contexts
Shows that models often use information better near the start or end of a prompt than in the middle.
Evaluation paper - 2023/2024
RAGAS: Automated Evaluation of Retrieval Augmented Generation
Defines reference-free RAG evaluation dimensions, including context relevance and answer faithfulness.
Research paper - 2023
MemGPT: Towards LLMs as Operating Systems
Uses a virtual context management metaphor to move information between memory tiers and the active prompt.
Cite This Page
Curated Context. "Context Engineering Glossary." Accessed July 6, 2026. https://curatedcontext.com/glossary
https://curatedcontext.com/glossary