自動検索されるLLMメモリと、明示的呼び出しのメモリの設計差ChatGPT and Gemini automatically inject past memories into unrelated…

Large language model providers have quietly diverged on one of the most consequential questions in conversational AI: how, and when, an assistant should recall what a user told it before. A recent analysis comparing ChatGPT, Gemini, and Claude concludes that the first two default to automatically retrieving and injecting past memories into new conversations, while Claude takes the opposite approach, waiting for an explicit request. The distinction matters because it shapes both the usefulness and the privacy profile of everyday interactions.

According to the write-up, the memory features in ChatGPT and Gemini are enabled by default and operate on an "automatic search, automatic injection" model. In practice, that means when a user opens a fresh chat to handle an unrelated task, the system may silently pull in fragments of earlier, sometimes personal conversations and weave them into the working context. The user is not necessarily told which memories are being surfaced, so personal details can shape a response even when the current request has nothing to do with them. The author reports that even an in-conversation reminder such as "let's keep this to work only this time" does not reliably prevent earlier context from being mixed in.

Claude, by contrast, reportedly uses an "explicit recall" design. Anthropic opened Claude's memory feature to all plans in March 2026, but structured it so that stored information is not injected into the context unless the user actively instructs the assistant to reference it. Under this model, memory behaves more like a filing cabinet the user chooses to open than an assistant that reaches for the file on its own. The trade-off is a more predictable context window at the expense of the seamless continuity that automatic injection can provide.

The comparison also flags a subtlety in Gemini's controls that is easy to miss. Turning memory off and deleting conversation history are two separate states. Disabling memory stops the feature from actively using stored details, but the underlying conversation logs appear to remain on the activity side of the account unless they are removed separately. Users who assume that switching memory off also erases their history may therefore retain more stored data than they expect, which is a meaningful consideration for anyone weighing what a provider holds on their behalf.

To understand why these choices differ, it helps to recall how memory is layered on top of a base model. A model's context window is finite, so persistent memory is typically implemented as a retrieval step that selects relevant stored snippets and prepends them to the prompt before generation, a pattern closely related to retrieval-augmented generation. Automatic injection maximizes convenience and personalization: the assistant appears to "know" the user across sessions without being asked. Explicit recall maximizes control and auditability, because nothing enters the context that the user did not request. Neither is inherently correct; they encode different assumptions about the default balance between helpfulness and restraint.

That tension has become more visible as memory has moved from an experimental toggle to a standard feature across the major assistants. OpenAI has expanded ChatGPT's memory over the past two years, and Google has extended personalization across its Gemini apps, while Anthropic's later entry gave it room to adopt a more conservative default. For users, the practical takeaways are straightforward. Anyone handling sensitive or compartmentalized work in ChatGPT or Gemini should assume that prior context can surface unless memory is disabled, and should verify whether history deletion is a separate action. Those who prefer strict separation between, say, personal and professional threads may find Claude's explicit model easier to reason about.

The broader point is that memory design is now a differentiator, not just a convenience. As assistants accumulate more about their users, the default behavior around retrieval determines how much of a person's history is silently in play at any given moment. The analysis stops short of declaring one approach superior, and the appropriate choice likely depends on whether a user values continuity or predictability. But it underscores that these defaults