Give Claude a Memory File for Long Projects

A pattern for work that spans many sessions: keep a single, hand-curated 'memory' document — key decisions, constraints, current state, open questions — and paste it (or attach it) at the start of each new chat. It's the durable spine the context window can't provide on its own.

A context window is per-conversation and finite; a long project has state that outlives any one chat. Without an external record you re-explain the same background every session and Claude re-derives decisions you already made — sometimes differently. A memory file makes state explicit and portable: each session starts from settled ground instead of from scratch, and the file itself becomes the source of truth you can review and correct. You control what persists, so drift and contradiction get pruned rather than accumulating.

Multi-week builds, ongoing research, anything you return to across days or many chats. Also whenever a single conversation gets long enough that you're about to start a fresh one and don't want to lose the thread.

One-off tasks that finish in a single session — a memory file is overhead with nothing to remember. And don't let it grow into a dumping ground; a bloated memory file has the same signal problem as any over-stuffed context.

Here is the project memory file — the durable state for this work:

<paste memory: goal, key decisions with rationale, current status, constraints, open questions>

Read it, then tell me in two lines where we left off and what the next decision is. At the end of this session, give me an updated version of this file with anything we settled or changed.

A three-week app build lives across a dozen chats. Each starts by pasting PROJECT.md — stack chosen, auth approach decided, schema so far, next feature. Claude picks up mid-stream instead of re-litigating the database choice, and ends each session by handing back an updated file. The document, not any single chat, is what carries the project.

Have Claude maintain the file for you: end each session with 'update the memory file — add what we decided, mark what's now done, flag anything that contradicts an earlier decision.' Keep sections stable (Decisions, State, Open) so diffs are easy to scan, and treat the file as reviewable — you approve what persists, exactly as you would a Claude Code CLAUDE.md.

• Letting the memory file grow unbounded until it's as noisy as the raw history it was meant to replace — prune it.
• Recording narration ('we discussed X') instead of decisions and current state, so it doesn't actually save re-explaining.
• Never reconciling contradictions, so old and new decisions coexist and Claude picks whichever it sees first.