Q
How long should CLAUDE.md be?
As short as honesty allows: only non-inferable constraints. If it exceeds a page, move workflows to on-demand skills and history to memory.
Glossary · CLAUDE.md
CLAUDE.md.
Claude Code's standing orders.
Direct answer
CLAUDE.md is the auto-loaded instruction file Claude Code reads at session start and preserves across compaction: project rules, architecture notes, commands. Treat it like code — versioned, reviewed, tested by observing behavior shifts — and keep it short, since bloated files get ignored and every line re-bills each turn. Durable facts live here; accumulating session knowledge belongs in a memory layer, not a growing markdown file.
In Anvaya
How We Implement It.
01CLAUDE.mdAnvaya respects CLAUDE.md as the Write layer and adds the memory layer beneath: weighted, decaying, drift-checked nodes served under budget.
Questions
Asked About CLAUDE.md.
Q
Why do agents ignore long instruction files?
Attention dilution plus lost-in-the-middle: rules sandwiched between file dumps read as background noise. Front-load, repeat critical constraints at the end.
Q
Does it survive compaction?
Yes — auto-loaded files persist across compaction, unlike one-off prompts. Another reason to keep them curated.
Stop Starting From Zero.
One binary. 11+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.