Search, Memory, and the State Around the Model

The product surfaces in the archive differ partly because they see different kinds of context. This chapter follows a question through current-date checks, retrieval, memory, preferences, workspace state, and connectors, so readers can distinguish what a model may already know from what a product supplies at request time and judge where each channel can fail.

“Memory” is often used as a single word for several different mechanisms. A product may carry the current conversation, retrieve an older chat, apply a saved preference, read a workspace file, consult a connector, or search the public web. These channels have different owners, different freshness, and different authority. The prompt captures are valuable because they sometimes describe the conditions under which each channel should be used.

Search is a decision about freshness

Search policy begins with a judgment about time. Claude’s captured instructions distinguish stable knowledge from information that can change, including current products, policies, positions, and prices. They direct the assistant to search when a remembered answer may be stale and to use connected internal sources for questions about a person’s or organization’s own data. Perplexity’s prompt provides explicit knowledge-cutoff and current-date context while presenting the assistant as a research-oriented guide. Gemini exposes search and retrieval tools alongside date-aware response guidance.

These rules define an epistemic threshold. The assistant is not simply asking, “Do I have a search tool?” It is asking, “Would answering from memory create a meaningful risk of being outdated, and which source has the right authority?” A stable mathematical definition may not need retrieval. A current product feature, officeholder, price, policy, or release does. A personal project question may need an internal connector rather than a public web search. The choice affects both accuracy and privacy.

Search introduces costs as well as benefits. It adds latency, can return irrelevant or conflicting material, and may expose a query to an external service. A prompt that directs search for every question may improve freshness while wasting time or increasing data exposure. A prompt that rarely searches may be fast while sounding confident about a world that has changed. The captured policies show products managing that trade-off through rules about rate of change, source type, and the user’s context.

Source choice is part of the answer’s authority. An internal drive or project record may be the right source for a user’s own document. A public source may be more appropriate for a general current event. A connector can hold privileged information that should not be mixed casually with public retrieval. The assistant’s role and the tool layer therefore meet at the source boundary. The prompt can prioritize one channel, but the host application and account permissions still determine what is actually accessible.

A returned result is not the same as verified knowledge. Search may produce a snippet, a page, a document, or a structured record. The assistant must decide what the result supports, whether sources conflict, and how much uncertainty to preserve. Claude’s captured search rules include source-handling and copyright constraints, which means retrieval and response composition are linked. Perplexity’s formatting rules similarly make structure part of the research experience. The product is specifying a path from query, to source, to explanation.

This path can fail in predictable ways.