The key is relevance, impact, and cost (token budget, noise). You should start by retrieving candidate context pieces (documents, memory entries) via embeddings, query matching, or metadata filters. Rank them by relevance (similarity, recency, importance). Then prune: pick the top-k that best contribute to answering the query without redundancy.
You also exclude conflicting or contradictory pieces, outdated information (context rot), or too-long texts that overwhelm the model. Summarizing or compressing long context pieces helps you keep key content without using all tokens. Use feedback loops: if you drop a piece and quality drops, reinstate it. Over time you learn which context types (history, memory, docs) matter most in your domain.
By selecting context that is semantically close, recent, non-redundant, and trustworthy, you allow the model to focus. Including too many or irrelevant pieces dilutes signal and increases hallucination risk. Therefore, context engineering is not just about adding more—it’s about carefully pruning and structuring the right set of context.