CLAUDE.md vs StremAI: two different memory jobs
They are complementary, and most developers should use both. CLAUDE.md and AGENTS.md hold instructions you write — conventions, build commands, rules — checked into the repo and read at session start. AGENTS.md made those instructions portable across many tools, which is genuinely solved. StremAI holds something different: memory agents accumulate while working — decisions, pitfalls, project context — stored as attributed entries and recalled selectively, so a session gets the few relevant memories rather than one ever-growing file. Keep your instruction files; StremAI does not replace them. It adds the learned layer they were never designed to carry.
Verified with Claude Code and Claude Desktop; Cursor, Codex, Windsurf, OpenClaw, and other MCP clients are supported or in active verification.
Markdown versionWhat CLAUDE.md is great at
CLAUDE.md and AGENTS.md are the right place for instructions you write: conventions, commands, repo policy, review rules, and team standards checked into source control.
Anthropic recommends adding to CLAUDE.md when Claude makes the same mistake again. That is a strong instruction workflow, and StremAI should not replace it.
The gap, as scope
Instruction files are authored by humans and loaded at session start. Claude Code auto memory is a learned layer, but Anthropic describes it as machine-local and not shared across machines or cloud environments.
StremAI is for learned context that should survive sessions and move across machines, repos, tools, and teammates, with selective recall instead of one growing file.
Using both
Keep CLAUDE.md for rules and commands. Add StremAI when agents discover decisions, pitfalls, deployment notes, or project context that another connected agent should recall later.
The split is simple: instructions are what you write; memory is what connected agents learn.
When you don't need StremAI
Skip it if you use one AI tool, in one repository, on one machine, and its built-in memory plus a CLAUDE.md or AGENTS.md file covers you; you work solo and are happy hand-curating notes into instruction files; or you want fully local, self-managed infrastructure. Open-source MCP memory servers are a reasonable choice if you prefer running your own. StremAI earns its place when agents span multiple tools, machines, or teammates, and when you want memory that is shared, attributed, and user-controlled without operating the layer yourself.
FAQ
Does StremAI replace CLAUDE.md or AGENTS.md?
They are complementary, and most developers should use both. CLAUDE.md and AGENTS.md hold instructions you write — conventions, build commands, rules — checked into the repo and read at session start. AGENTS.md made those instructions portable across many tools, which is genuinely solved. StremAI holds something different: memory agents accumulate while working — decisions, pitfalls, project context — stored as attributed entries and recalled selectively, so a session gets the few relevant memories rather than one ever-growing file. Keep your instruction files; StremAI does not replace them. It adds the learned layer they were never designed to carry.
What should go in CLAUDE.md?
Put explicit instructions there: commands, conventions, repo rules, coding standards, and review policy.
What should go in StremAI?
Store learned memory: decisions, pitfalls, architecture notes, handoffs, and recurring context agents discover while working.
How do I make Claude Code remember things across machines?
Keep CLAUDE.md for shared instructions, then connect Claude Code to a shared MCP memory layer like StremAI so learned entries can be recalled from other machines and sessions.
How do I share Claude Code learning with my team?
Connect team agents to the same StremAI project. What one connected agent stores as a useful lesson can be recalled later by teammates with access.
What is the difference between instructions and memory for AI agents?
Instructions are rules humans write before work starts. Memory is context agents learn while working, such as decisions, pitfalls, and handoffs.
Sources
- Anthropic Claude Code memory docs accessed 2026-07-07 — Instruction files and machine-local auto memory scoping.
- AGENTS.md accessed 2026-07-07 — Portable instruction-file convention.
Start with a real connection
OAuth/browser sign-in is preferred. API keys stay available for CI, scripts, and clients that cannot complete OAuth.