Auto-memory should allow recall + user-requested writes while disabling autonomous mid-session writes

Type: Feature request (+ minor docs gap)

Problem

autoMemoryEnabled is binary. Per the docs, when false "Claude does not read from or write to the auto memory directory." There's no way to keep memory recall (reading existing memories across sessions) while stopping Claude from autonomously pausing mid-session to write new, often low-context memories.

A user who finds the auto-authored memories terse/low-quality must disable the whole feature — losing recall too, which defeats the point of memory.

Desired behavior

Decouple read from write, or add a memory mode:

• read-only (recall yes, no writes), or
• confirm-on-write (every memory write prompts the user), or
• manual (memories created only when the user explicitly asks, never on Claude's own initiative, with recall still on).

Net goal: read = yes, user-requested write = yes, autonomous/proactive write = no.

Why the current options don't cover it

• autoMemoryEnabled: true → recall + autonomous writes (the unwanted behavior).
• autoMemoryEnabled: false → no writes, but also no recall.
• CLAUDE.md guidance ("don't self-author memories") is only an in-context instruction, not enforced — the model can still slip.
• A PreToolUse hook on the memory directory can enforce confirm/deny, but it cannot distinguish a user-requested write from an autonomous one, and is non-trivial to author. This really wants a first-class setting.

Docs gap (minor)

• autoDreamEnabled appears in the settings JSON schema (background "memory consolidation") but is undocumented at code.claude.com/docs; its semantics and relationship to mid-session writes are unclear.
• The trigger for when Claude decides to write a memory mid-session isn't documented, so users can't reason about or shape the behavior.

Environment

Claude Code, model Opus 4.8 (1M context). Refs: settings docs, memory docs

🤖 Generated with Claude Code