Retrieval searches one or more information stores and chooses records likely to help with a question or decision. It may use keywords, semantic similarity, metadata, graph relationships, structured queries, or combinations of these methods.
Good retrieval balances relevance, freshness, authority, and access control. Returning too little can hide crucial facts, while returning too much or conflicting material can waste context and make the agent's reasoning less reliable.
ELI5
Retrieval is finding the useful notes before starting the next step. The memory store may contain many records, but the agent should receive only the ones that matter for its current job.
For example, when a developer resumes a branch, retrieval might return the branch goal, the latest decision, and the last failing test. It should not mix in unrelated conversations from another project.
