Loading...
From first setup to self-improving. Build an agent that remembers permanently — not within a chat, but across weeks. Based on official Anthropic docs + the Dreaming research preview.
Based on research by 0xCodez
The 4 layers
Layer 1: Sticky Note
Instructions that persist inside a Project. Lost when you leave.
Layer 2: Notebook
Chat Memory + seeded preferences. Survives across all conversations.
Layer 3: Filing Cabinet
Structured memory files (CLAUDE.md, memory.md). Survives across sessions and projects.
Layer 4: Self-Improving
Dreaming. The agent reviews its own sessions and rewrites memory to get sharper every week.
What most people skip. Takes 10 minutes.
Claude has persistent memory since March 2026. It remembers preferences, projects, and working style across conversations automatically.
Profile icon → Settings → Capabilities → Memory section → confirm "Generate memory from chat history" is on.
Don't wait for Claude to infer preferences (synthesis runs ~once/24h). Tell it directly. Explicit beats inferred, and it lands immediately.
In a fresh chat: "Remember the following about me: I work in [field], prefer [style], never [the thing you always correct]."
A Project is a persistent workspace where instructions stay loaded across every conversation inside it. Name it after the job, not the topic.
Projects → New Project. Fill custom instructions with agent role, standards, and constraints.
Projects persist instructions, NOT conversation history. New chat in the same Project = all discussion history gone. This burns most people.
Accept this limitation. Build Layer 3 (memory files) to bridge the gap.
For Claude Code users. The memory that survives across sessions.
A single file the agent reads at start and appends to at end. In Claude Code this is CLAUDE.md. Keep it lean — ~20K tokens max before you type anything.
Create .claude/memory.md with sections: Preferences, Decisions (dated), Known workarounds, Recurring mistakes.
Claude Code writes notes to itself based on your corrections. Corrections survive into the next session instead of evaporating.
Toggle with /memory in a session, or set autoMemoryEnabled in project settings.
A memory file that grows without structure becomes noise. Each entry earns its place. This is the file the agent consults to stop repeating yesterday's mistakes.
Organize into: Preferences, Decisions (dated), Known workarounds, Recurring mistakes to avoid.
Not everything should be saved. Would this change how the agent acts next time? If yes, store it. If no, let it go. Memory that stores everything is as useless as memory that stores nothing.
After each session: extract only decisions, workarounds, preferences, and failure modes.
The Dreaming API (May 2026). Agents that get smarter every week.
Shipped May 2026 as a research preview. A scheduled background process that reads past sessions and produces a reorganized memory store: duplicates merged, stale entries replaced, new insights surfaced.
Only for agents running the same task repeatedly. A one-off agent has nothing to consolidate.
Requires Managed Agents API key + Dreaming access. Pass existing memory store + up to 100 session IDs. Dreams run on Opus 4.7 or Sonnet 4.6.
client.beta.dreams.create() with memory_store + session_ids. Start with small batch, scale once quality is verified.
The dream produces a SEPARATE output store. Input stays read-only. Review that merged entries are correct and surfaced insights are real. This is the safety net.
Read the output store. Check merged entries, replaced entries, and new insights. Never auto-deploy.
Point agent at the new store. Schedule dreaming nightly or weekly. The agent works during the day, dreams between runs, comes back sharper each cycle.
Harvey (legal AI) reported 6x increase in task completion after enabling Dreaming.
Treating Projects as memory
Projects persist instructions, not conversation history.
Dumping everything into CLAUDE.md
Lean and structured beats long and complete. ~20K token budget.
Storing with no filter
If everything is worth remembering, nothing is. Save only what changes future behavior.
Auto-deploying dream output
Always review the output store before deploying. That's the safety net.
Dreaming on low-frequency agents
Only consolidates patterns across many sessions. Twice a month = nothing to consolidate.
7-tier memory architecture, structured memory.md templates, HALO self-improvement loop, and the complete brain sync protocol. Install in 30 seconds.
Get AI Brain Pro — $97