Anthropic Auto Dream
https://claudefa.st/blog/guide/mechanics/auto-dream
Background sub-agent in Claude Code that periodically scans ~/.claude/projects/<project>/memory/ + recent JSONL transcripts via grep, prunes stale notes, deduplicates, resolves contradictions, converts relative dates to absolute timestamps, and rewrites MEMORY.md (kept under 200 lines). Triggers automatically when 24+ hours elapsed AND 5+ sessions completed; manually via /dream.
At a glance
- Type
- REM-style sub-agent memory consolidation
- Tier
- T1
- Section
- Claude Code memory mechanisms
- Created
- 2026-05 (research-preview rollout in Claude Code; server-side feature flag)
- Latest release
- not applicable — not OSS (server-side feature flag)
- License
- not applicable — not OSS
- GitHub
- not applicable — no GitHub repo (Claude Code closed source)
- Pricing
- bundled with Claude Code (no separate fee)
- Funding
- Anthropic total $12.4B+ raised; $40B valuation
Taxonomy
- storage
- file
- retrieval
- extraction-pull
- persistence
- long-term
- update
- consolidation
- unit
- fact
- governance
- user-controllable
- conflict
- llm-arbitrate
When to use
Optimised for: background memory consolidation + bounded MEMORY.md size + surgical updates
Anti-fit: not for users on Claude.ai web (CLI-only); not for environments where consolidation logic must be auditable
Pros & cons
Pros
Bounded MEMORY.md size (200 lines) prevents context bloat; surgical updates preserve user-edited content; user-toggleable; complements rather than replaces CLAUDE.md.
Cons
Closed-source; opt-out mechanism not formally documented; consolidation logic opaque; no audit log for what was pruned/merged; only Claude Code (not Claude.ai web).
Claims & capabilities
Consolidation of 913 sessions in ~8-9 minutes; surgical updates only modify changed files; read-only mode prevents source-code edits; lock file prevents concurrent runs (vendor-reported)
Technical surface
- API surface
- /dream slash command + /memory toggle (Auto-dream: on)
- Backend storage
- local filesystem (~/.claude/projects/<project>/memory/)
- Deployment
- Local CLI (Claude Code) — Claude.ai web not in scope
- Embedding model
- not applicable — files are markdown (no embedding step)
- Multi-tenancy
- not applicable — single-user local
- MCP
- native (first-party — Claude Code built on MCP)
- A2A
- not supported
- OpenTelemetry
- no — local CLI feature
Compare Anthropic Auto Dream with…
Similar systems
Other claude code memory mechanisms in the catalog, ranked by inbound references.
- Official MCP Memory server T1
Knowledge-graph-based persistent memory MCP server — the canonical reference implementation. npx -y @modelcontextprotocol/server-memory ; configurable via MEMORY_FILE_PATH .
- Claude Code auto-memory T1
Persistent file-based memory at ~/.claude/projects/<project>/memory/ with a MEMORY.md index plus typed memory files (user / feedback / project / reference). Loaded into context across sessions.
- claude-brain T2
Claude Code plugin packaging of Memvid (see Dedicated memory layers). Wraps the Memvid .mv2 single-file store as a slash-command-installable plugin so per-project memory becomes a git-committable file inside the repo.
- Claude-CursorMemoryMCP T2
FastAPI server. RBAC + API-key auth, encryption, backup/recovery, embedding pipeline with deduplication. Only server in this list with role-based access control over pgvector.
- claude-mem T1
Hooks into Claude Code's session lifecycle; compresses tool outputs into ~500-token observations via Claude's agent SDK. SQLite + FTS5 with optional Chroma vector. "Endless Mode" compresses in-flight to extend session length. Captures what the agent did , not just what the user typed.
- claude-supermemory T2
Wraps Supermemory cloud API. Team memory shared across project members, separate from personal memories. Auto-capture on session end; per-repo config.
Related systems
References (1)
- Anthropic Claude (foundation models) depends on at runtime — adjacent-infrastructure cell: requires Claude Code v2.1.59+; local filesystem; no external store