Claude user reports better responses with memory disabled

The user, a Claude subscriber on the $20/month plan, has been using the model for research, rubber duckying (testing ideas by explaining them aloud), and as a creative soundboard. Recently they noticed a significant decline in response quality when working in high-context situations with Opus 4.8. In an effort to improve performance, they disabled Claude's memory system, which is now labeled "Legacy" despite the absence of an announced replacement feature.

After disabling memory, the user observed a drastic improvement in response quality and accuracy. Although this required them to manually re-input necessary context for task continuations, the trade-off proved worthwhile. The user theorizes that a change introduced around Opus 4.7 — likely designed to expand the memory context window — may have inadvertently reduced actual response quality. To test this theory, they point to a research task spanning upwards of 30 prompts: when they switched to a fresh chat and condensed the discussion into a single new prompt, the results were superior to continuing the original conversation, despite (or perhaps because of) the loss of the prior chat's full context.

The user also raises a broader question: whether the high-context abilities of state-of-the-art thinking models are genuinely as strong as claimed, or whether models perform better with carefully filtered context than with exhaustive conversation history.

A secondary frustration emerges around feature granularity. The user wishes that "project" and "chat" memory could be controlled independently. Disabling memory globally boosts chat capability but cripples projects by removing their internal memory, forcing the model to search for files each time a reference is requested. The user acknowledges that creating a project for every task might be a workaround, but dismisses it as tedious.

The post reflects a user's empirical observation that Claude's memory feature, introduced or modified in recent versions, may introduce a quality trade-off in high-context scenarios. The user's theory centers on a hypothesis: that changes made around Opus 4.7 or 4.8 to expand memory context window inadvertently degraded response coherence or accuracy, possibly by diluting model focus or introducing interference from accumulated context. The anecdote of the 30-prompt research task is revealing — rather than memory enriching the conversation, the user found that starting fresh with a condensed summary produced better results, suggesting that the model may perform better with filtered, curated context than with the full conversational history.

The user's secondary complaint — that memory settings apply globally rather than per-feature — highlights a product design tension. Anthropic appears to bundle memory control together, but the user's workflow reveals that chat and project contexts have different needs: chat benefits from memory being off, while projects require it for file referencing. This suggests the settings granularity may not yet match real user workflows.

No comments yet. Be the first to share your thoughts!