How Claude Code’s Upcoming Auto‑Dream Feature Organizes Memory Like Human Sleep

According to reports on Reddit, Claude Code has quietly enabled a not‑yet‑public feature called Auto‑dream under the /memory settings, currently showing the status "off never".

The feature’s logic is to run a sub‑agent on a regular schedule that scans Claude’s short‑term context, extracts high‑value information, archives it to long‑term storage, and optionally prunes redundant data.

Cursor already offers a similar capability via its Continual Learning plugin, which uses an agents‑memory‑updater sub‑agent to mine valuable details from conversation logs, update the AGENTS.md file, and keep an incremental index synchronized.

User feedback on Cursor’s plugin notes that it mainly consumes tokens, often triggering at the end of a conversation and wasting the remaining context window.

Developers of AI coding assistants have been grappling with the problem of retaining useful information over long‑running projects without overloading the context window. Manually maintaining memory files is labor‑intensive and easy to forget, so both Auto‑dream and Continual Learning aim to automate this process.

Claude Code’s designers call the mechanism "dream" because, like the human brain during sleep, the AI replays daytime inputs, solidifies important fragments into long‑term memory, and trims irrelevant details.

A practical challenge discussed in the community is deciding what to retain: overly aggressive deletion may discard critical business decisions, while overly conservative retention fills the memory file with noisy debugging data.

Although Auto‑dream is not yet publicly available and Claude Code’s memory functions remain isolated per client, the direction indicates a shift from merely stuffing more data into AI assistants toward smarter memory management.