CLAUDE.md vs Claude Projects vs MCP: What Goes Where (2026)
Direct answer: Put repository build commands and coding rules in CLAUDE.md, web workspace briefs and reference PDFs in Claude Projects, and cross-tool conversations or settled decisions in a remote MCP vault like Multiplist when work spans Chat, Code, and ChatGPT.
You hammer out an API design inside a Claude Project in your browser, then switch to Claude Code in your terminal and realize it has no idea what you just settled. Meanwhile, your CLAUDE.md file keeps swelling with pasted planning notes that only mattered for one sprint.
Anthropic gives you three different context layers in 2026, and each solves a different job. Knowing where to put instructions, files, and past reasoning stops context bloat before it starts.
Add Multiplist to Claude or unpack an existing export at rescue.multiplist.ai.
Official Claude Code Memory Docs, Claude Code MCP Docs, Claude Projects Guide, Project Retrieval Docs, and Multiplist docs checked November 21, 2026.
# Side-by-Side: CLAUDE.md vs Claude Projects vs MCP
| Dimension | CLAUDE.md (+ Auto Memory) | Claude Projects | Remote MCP (Multiplist) |
|---|---|---|---|
| Primary home | Local git repo & ~/.claude/ | claude.ai & Claude Desktop | Hosted vault (https://multiplist.ai/mcp) |
| Best suited for | Build commands, repo rules, style | Web briefs, PDFs, team docs | Saved conversations & cross-AI decisions |
| Who writes it | You (CLAUDE.md) / Claude (MEMORY.md) | You (files/rules) & Claude (project memory) | You choose what to save or extract |
| Loaded into | Every Claude Code session in scope | Chats inside that Claude Project | Any authorized Claude, Code, or ChatGPT chat |
| Practical limits | Target <200 lines; MEMORY.md caps at 200 lines / 25KB | Context window + automatic Project RAG | Deliberate save & retrieval per source |
| Exact wording | Verbatim file text in repo | Uploaded files + summarized memory | Original conversation text + line citations |
| Cross-AI reach | Claude Code (or AGENTS.md tools) | Claude web & Desktop only | Claude Chat, Cowork, Code & ChatGPT |
# What Belongs in CLAUDE.md (and Auto Memory)
Per Anthropic's Claude Code memory documentation, Claude Code splits persistent repository context into human-written CLAUDE.md files (or AGENTS.md files) and Claude-written auto memory:
CLAUDE.mdfiles: Plain-text Markdown (./CLAUDE.md,./.claude/CLAUDE.md,~/.claude/CLAUDE.md, or org-level policy) read at the start of every session. Anthropic recommends keeping each file under 200 lines with concrete, verifiable rules—like test commands, folder layout, and style conventions—and moving path-specific rules into.claude/rules/*.md.- Auto memory (
MEMORY.md): Notes Claude writes itself per repository (shared across git worktrees) to remember corrections, build quirks, and preferences. Only the first 200 lines or 25KB ofMEMORY.mdload automatically at session start.
Strengths and limits: Version-controlled CLAUDE.md keeps an entire engineering team aligned inside the codebase. Still, both layers act as context rather than strict enforcement (use PreToolUse hooks to block tool calls reliably), and Claude web Chat never sees your local CLAUDE.md automatically.
# What Belongs in Claude Projects
Per Anthropic's Claude Projects and Project Retrieval guides, a Claude Project bundles three layers inside your web or desktop workspace:
- Project Instructions: Standing directions for tone, output structure, or workflow rules across every conversation in that workspace.
- Project Knowledge: Uploaded specs, briefs, and PDFs. When your knowledge files approach the context window limit, Claude switches to retrieval mode so relevant passages are pulled on demand.
- Project Memory: Separate workspace memory that evolves with your project chats, with a Use account memory toggle when you want a workspace isolated.
Strengths and limits: Projects work well for strategy, writing, and multi-document research inside Claude. However, project files stay inside claude.ai: Claude Code in your terminal cannot query your web Project Knowledge directly, and ChatGPT cannot read it either.
# When to Add a Remote MCP Vault Like Multiplist
Use a remote Model Context Protocol (MCP) vault when useful reasoning happens in a conversation and needs to follow you across Claude Chat, Claude Code, and ChatGPT—without bloating CLAUDE.md or re-uploading files by hand.
- Save the conversation or decision. Capture a live Claude or ChatGPT thread in one click with the free Multiplist Chrome Extension, unpack a bulk
conversations.jsonexport at rescue.multiplist.ai (a hosted upload processed in server memory), or save a note directly. - Connect Multiplist to each surface. Click Add Multiplist to Claude for Chat and Cowork, or register the same remote endpoint in Claude Code (
claude mcp add --scope user --transport http multiplist https://multiplist.ai/mcp, thenclaude mcp login). - Retrieve the exact passage. Ask Claude Code or ChatGPT to pull the named decision, framework, or insight—keeping the original conversation attached with line citations back to what you actually said.
Boundary: Connecting Multiplist never automatically collects your past conversations or watches your local CLAUDE.md. You decide what to save, enable the connector where you work, and inspect returned sources. Hosted vault storage uses Multiplist's service (privacy policy).
# A Simple Rule for Where Context Goes
- Repository rules and test commands: Put them in
CLAUDE.md(under 200 lines) or.claude/rules/. - Multi-document specs for browser strategy: Upload them to a Claude Project.
- Settled decisions and conversations needed across Claude Chat, Claude Code, and ChatGPT: Save them in a Multiplist vault over remote MCP—starting free on the onboarding page.
Related: ChatGPT Projects vs Claude Projects (2026), Claude Memory vs ChatGPT Memory (2026), Link Claude Code Terminal Sessions with Claude Chat, and Best MCP Servers for AI Memory (2026).