Auto-Memory and Auto-Dream — what Claude Code does at night
Anthropic added Auto-Memory in March 2026 and Auto-Dream consolidates it nightly. Useful by default. Dangerous when combined with external memory systems. Here is how to keep both happy.
Auto-Memory and Auto-Dream — what Claude Code does at night
Since March 2026, Claude Code has had two memory features that are default-on and that most users do not realize are running. This recipe explains what they do, when they help, and the one situation where they cause real problems.
Step 1: Understand what each feature actually does
Auto-Memory. While you work, Claude writes notes in the background to a local memory file. Build commands, architecture decisions, your preferences, things you correct it on. Across sessions these accumulate.
Auto-Dream. A background subagent that runs once every 24 hours across all memory files. It deduplicates, removes outdated info, and rewrites relative dates ("yesterday") into absolute dates ("on 2026-03-15") so the file does not read confused six months from now. Anthropic markets it as the "nighttime brain" of the AI.
The settings live in ~/.claude/settings.json:
{
"autoMemoryEnabled": true,
"autoDreamEnabled": true,
"cleanupPeriodDays": 30,
"autoCompactWindow": 200000,
"fileCheckpointingEnabled": true
}
autoMemoryEnabled is true by default. autoDreamEnabled rolls out via a server-side feature flag — you may or may not have it active yet. Check with /memory in a Claude Code session — the selector should show "Auto-dream: on" if it is active for your account.
Step 2: Identify if you are in the happy case
You only use Claude Code. You have no other memory system. You are happy with the AI writing into ~/.claude/CLAUDE.md (the user-scope file) and per-project CLAUDE.md files.
In this case: leave both default-on. Check weekly what got written. If the file accumulates noise, you can prune by hand, or disable for a day, or run /memory and edit directly.
Step 3: Know the three structural limitations
Even in the happy case, three things are worth knowing.
It is local and Claude-Code-only. If you also use Cursor, Codex, or any other AI tool, those tools see none of what Claude Code wrote. Auto-Memory has no cross-tool story.
It is markdown-file-based. No knowledge graph, no semantic retrieval, no confidence score, no bi-temporal model. If your CLAUDE.md grows past a few hundred lines, the model starts missing things — there is no smart query layer, just full-file read.
Auto-Dream's nightly run can collide with external memory systems. If you also run a memory MCP that writes to ~/.claude/, Auto-Dream might consolidate or scramble those files at 2 AM. This is the problem this recipe really exists to warn you about.
Step 4: Pick a collision-avoidance strategy
If you run an external memory system (StudioMeyer Memory, local-memory-mcp, Mem0 self-hosted, anything that writes to ~/.claude/ for hooks, settings, or memory files), Auto-Dream can mistake your files for its own. Three options to handle this.
Option A: Turn off Auto-Memory and Auto-Dream, let your external system be master.
Recommended if you have invested time in a real memory MCP setup (Recipe 4.1).
Edit ~/.claude/settings.json:
{
"autoMemoryEnabled": false,
"autoDreamEnabled": false
}
Now Claude Code does no automatic memory writes. Your MCP server is the single source of truth. The always-on instruction block (Recipe 4.1) tells the model when to call nex_search, nex_learn, etc.
Option B: Move external memory files outside ~/.claude/.
If you want both systems but keep them separate, store your external memory's local files anywhere except ~/.claude/. For example: ~/.studiomeyer-memory/, ~/Library/Application Support/local-memory-mcp/, ~/Documents/memory/. Auto-Dream only touches ~/.claude/.
Option C: Carefully separate by file naming inside ~/.claude/.
If you must store inside ~/.claude/ (because of plugin paths or similar constraints), Auto-Dream targets specific filename patterns. Your external system's files should not match CLAUDE*.md, memory.md, or be in ~/.claude/agents/{agent-name}/memory/. Use a non-conflicting subdirectory like ~/.claude/external-memory/ and your system's tools will not interfere.
Of the three, Option A is the cleanest. Option B is the least invasive if you actually use Auto-Memory for non-MCP-related notes. Option C is the workaround for unusual setups.
Step 5: Verify which mode you are in
Quick test: is Auto-Dream actually running for you?
Open Claude Code. Run /memory. The selector at the top right should show one of:
- "Auto-dream: on" — feature flag enabled, runs every 24h
- "Auto-dream: off" — feature flag not enabled yet, or you disabled it
- (no line) — your version is older than the rollout
You can also trigger it manually by saying "consolidate my memory files" — Claude will run the same logic on demand. Useful for testing without waiting 24 hours.
What you might lose by turning Auto-Memory off
Honest disclosure. Auto-Memory captures the small things — a build command you found, a quick fix you discovered, the exact npm script that works. Those land in CLAUDE.md and are searchable next session.
If you turn it off, you have to write those manually — either via your MCP server's nex_learn tool (or equivalent) or by editing CLAUDE.md by hand.
In our experience the trade is worth it. Auto-Memory is fine for individuals who only use Claude Code. The moment you have a real cross-tool memory setup, the noise of two systems writing into the same files is worse than the convenience of one system writing automatically.
What's next
If you are running a memory MCP server, you have made the call about Auto-Memory. Recipe 4.8 compares the major server options head-to-head with honest benchmark numbers. Recipe 4.9 covers the failure modes that get worse as your memory accumulates — particularly sycophancy amplification and stale facts.