Preflight Checklist
What's Wrong?
Claude Code (via Claude Desktop, Windows) writes automatic memory files to /memory/ (e.g. MEMORY.md plus per-topic files) even though every memory-related setting I can find is switched off.
What Should Happen?
With memory settings off, no memory files should be written and the auto-memory system prompt instructions should not be injected.
Error Messages/Logs
Steps to Reproduce
Steps to reproduce
- Open Claude Desktop → Settings → Capabilities → Memory.
- Confirm both "Search and reference chats" and "Generate memory from chat history (Legacy)" are toggled off.
- Toggle both on, then back off again, to rule out a stale state.
- Start/continue a Claude Code session in a project.
- Observe that Claude still writes/updates files under .claude/projects//memory/ (MEMORY.md index plus topic files) during the session, and the system prompt still contains the full "auto memory" instruction block telling it to do so.
Claude Model
No response
Is this a regression?
I don't know
Last Working Version
No response
Claude Code Version
2.1.193 (Claude Code)
Platform
Other
Operating System
Windows
Terminal/Shell
Other
Additional Information
Memory files continue to be created/updated.
No toggle exists anywhere in Settings (General, Capabilities, Claude Code section) that maps to this specific file-based project-memory feature — the two toggles under Capabilities → Memory appear to control claude.ai's own chat-history memory, not Claude Code's project memory system.
Checked ~/.claude/settings.json, project .claude/settings.json/settings.local.json — no relevant key found.
Checked cached feature flags in ~/.claude.json and found "tengu_session_memory": false — suggesting the client believes memory is disabled, yet the behavior contradicts that. This looks related to #77627, which reports the same flag not gating memory tool injection (there manifesting as "tool not found" errors on macOS/iTerm). In my case, on Windows/Claude Desktop, there are no errors — the memory tools work normally and silently save files, so the flag appears to be ignored rather than causing a partial/broken gate.
Preflight Checklist
What's Wrong?
Claude Code (via Claude Desktop, Windows) writes automatic memory files to /memory/ (e.g. MEMORY.md plus per-topic files) even though every memory-related setting I can find is switched off.
What Should Happen?
With memory settings off, no memory files should be written and the auto-memory system prompt instructions should not be injected.
Error Messages/Logs
Steps to Reproduce
Steps to reproduce
Claude Model
No response
Is this a regression?
I don't know
Last Working Version
No response
Claude Code Version
2.1.193 (Claude Code)
Platform
Other
Operating System
Windows
Terminal/Shell
Other
Additional Information
Memory files continue to be created/updated.
No toggle exists anywhere in Settings (General, Capabilities, Claude Code section) that maps to this specific file-based project-memory feature — the two toggles under Capabilities → Memory appear to control claude.ai's own chat-history memory, not Claude Code's project memory system.
Checked ~/.claude/settings.json, project .claude/settings.json/settings.local.json — no relevant key found.
Checked cached feature flags in ~/.claude.json and found "tengu_session_memory": false — suggesting the client believes memory is disabled, yet the behavior contradicts that. This looks related to #77627, which reports the same flag not gating memory tool injection (there manifesting as "tool not found" errors on macOS/iTerm). In my case, on Windows/Claude Desktop, there are no errors — the memory tools work normally and silently save files, so the flag appears to be ignored rather than causing a partial/broken gate.