How to Preserve Project Context Across Claude Code Sessions

Claude Code starts each session with a fresh context window, the information it can use while working in that conversation. Save project guidance that applies to future sessions in CLAUDE.md, so you do not have to repeat it. To continue a task that depends on an earlier discussion, resume that session or give the next session a short handoff.

Each option carries forward a different kind of context: a project instruction file gives Claude reusable guidance, auto memory saves selected learnings on one machine, and resuming a session restores its conversation. Choose the option that fits what you need to carry forward, then check which files Claude loaded.

Put lasting project guidance in CLAUDE.md

A project’s CLAUDE.md file gives Claude instructions that apply across sessions. Add details that are useful for future work, such as build and test commands, project conventions, architecture decisions, and common workflows. Anthropic recommends keeping each file under 200 lines and writing concise, concrete instructions in its memory documentation.

You can place project instructions in ./CLAUDE.md or ./.claude/CLAUDE.md and commit them so they are available to other contributors. Use ~/.claude/CLAUDE.md for instructions that apply to your own work across projects. For personal guidance limited to one project, use CLAUDE.local.md and add it to .gitignore so it stays out of version control.

From the repository, run /init to create a starting CLAUDE.md based on what Claude finds in the code. Review the result: add decisions or details that the code does not reveal, and remove instructions that are obvious, outdated, or irrelevant to most tasks. The Claude Code best practices recommend recording information Claude would otherwise have to rediscover.

If a rule applies only to part of the codebase, keep it out of the project-wide file. Put guidance for specific file paths in .claude/rules/. Put a multi-step procedure in a skill if Claude needs it only for certain tasks. This keeps general instructions focused and limits unrelated context.

Treat auto memory as a personal supplement

Claude Code’s auto memory lets Claude save selected project learnings, preferences, and corrections. It is stored locally under ~/.claude/projects/<project>/memory/, so it does not serve as a shared record across machines or cloud environments. Use a committed project file or another shared project document for information that must be available to other contributors or on another device.

Claude automatically reads only the first 200 lines or 25 KB of MEMORY.md, whichever comes first. Other memory files are available to read when needed, but they do not all load at session start. Use /memory to review or edit memory files, and /context all to check which files Claude loaded. These details and the documented limits are covered in the memory documentation.

Resume, compact, or start a fresh session

If you need details from an earlier conversation, use claude --continue to reopen the most recent session in the current directory, or claude --resume to choose a saved session. You can also run /resume from an active session or name a session with /rename so you can find it later. Claude Code saves CLI sessions locally, and resuming restores conversation history, tool calls, and results, as described in the session management documentation.

If you are continuing the same task after a long conversation, run /compact to have Claude summarize the history so work can continue with less conversation context. You can give /compact instructions about what the summary should retain, such as decisions made or tests still to run. A summary is not a full transcript, so check that the details needed for the next step remain available.

For a genuinely new task, use /clear to start a fresh conversation. If work is unfinished, give the new session a brief handoff. State the goal and what you have completed, then note what remains. Include important decisions, relevant files, and the next command or check. Keep enduring project knowledge in CLAUDE.md; reserve the handoff for details specific to the current task.

Claude does not automatically restore some startup options when you resume. Anthropic notes that options such as --settings, --mcp-config, and --plugin-dir, along with directories added using --add-dir, may need to be supplied again. If a resumed conversation lacks a tool or directory it used before, check which startup options and directories your session needs.

Check instructions and remember their limits

After you edit project guidance, run /context all to confirm that Claude loaded the expected memory files. Run /memory to inspect the memory Claude can use. If Claude misses an instruction, check whether it applied to the task and whether Claude loaded it. Then make the instruction more specific or move it to the right file.

CLAUDE.md and auto memory provide context that guides Claude’s responses. They do not enforce rules on every turn. Use a mechanism such as hooks, deny rules, or managed settings when a rule must be enforced, rather than relying on remembered instructions alone.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com