How to give your Claude agent a memory in 12 steps: from first setup to self-improving

A practical walkthrough for giving Claude agents persistent memory that survives across sessions and improves over time. The guide builds four layers of increasing sophistication, from built-in features to Anthropic's Dreaming research preview.

Foundation (Layers 1–2). Claude's built-in Chat Memory, rolled out March 2026, synthesizes preferences from conversation history roughly every 24 hours. The author recommends seeding it explicitly rather than waiting — one deliberate message lands immediately. Projects provide persistent instructions across chats inside the workspace, but do not retain conversation history by default; that gap is where people get burned.

Persistent memory for coders (Layer 3). A living memory file (CLAUDE.md in Claude Code) gives the agent a structured record of preferences, decisions, workarounds, and recurring mistakes. Auto-memory in Claude Code writes back corrections automatically. The key discipline: save only what would change future behavior. A memory file that stores everything is as useless as one that stores nothing.

Self-improving memory — Dreaming (Layer 4). Shipped as a research preview in May 2026, Dreaming is a scheduled background process that consolidates session transcripts and existing memory into a new, reorganized memory store: duplicates merged, stale entries replaced, new insights surfaced. It requires a Managed Agents API key and gated access. The dream produces a separate output store so it can never corrupt the original; the author stresses reviewing output before swapping it in. Harvey, a legal-AI company, reported roughly a 6× increase in agent task-completion rates after enabling Dreaming for legal-drafting workflows.

The article closes with four common mistakes: treating Projects as memory, bloating CLAUDE.md, storing everything with no filter, and auto-deploying dream output without review.