The original title is "CLAUDE.md vs. a memory MCP: what actually goes where"

Short summary

CLAUDE.md holds stable, human-curated facts loaded into context every session, while a memory MCP server stores dynamic facts the agent accumulates during work and queries on demand. The author provides a practical framework: if you'd type it by hand, it belongs in CLAUDE.md; if it's a byproduct of doing the work, it wants a memory store. Four triggers—multiple tools, multiple machines, team-shared context, and volume past hand-curation—justify adding a memory server.

• •CLAUDE.md is for stable, human-curated facts; memory MCP is for dynamic facts the agent accumulates while working
• •The test: would you type this by hand? If yes, CLAUDE.md; if it's a byproduct of work, use a memory store
• •Four triggers for adding a memory server: multiple tools, multiple machines, team-shared context, and volume past hand-curation

Generated with AI, which can make mistakes.