CLAUDE.md
https://docs.claude.com/en/docs/claude-code/memory
Markdown file injected at session start; reloaded post-compaction. Document-as-memory: human-readable, version-controlled, no extraction or vector store. Anthropic-canonical pattern; equivalent emerges in Cursor / Windsurf.
At a glance
- Type
- Git-backed persistent instructions
- Tier
- T1
- Section
- File-backed / editor paradigms
- Created
- 2024-10 (first appeared in Claude Code public release; Claude Code beta May 2024; GA Oct 2024)
- Latest release
- not applicable — not OSS
- License
- not applicable — not OSS
- GitHub
- not applicable — no GitHub repo
- Pricing
- not applicable — not commercial
- Funding
- not applicable — not commercial
Taxonomy
- storage
- file
- retrieval
- injection
- persistence
- cross-session
- update
- overwrite
- unit
- document
- governance
- user-controllable
- conflict
- manual
When to use
Optimised for: simplicity / git-versioned / human-readable
Anti-fit: not for non-coding / non-IDE use cases
Pros & cons
Pros
Plain-text + git-versioned + human-reviewable — the highest-trust memory format with zero infra; works in any editor or terminal.
Cons
Manual curation required; doesn't scale beyond ~10–20 KB before context bloat; no semantic recall, only injection-by-default.
Claims & capabilities
GitHub Copilot leads at 41.8% market share (2026); Cursor 27.3%; Claude Code 12.5% within 12 months of release; Windsurf 9.4%. CLAUDE.md is the canonical config-file pattern that Claude Code reads on every session for project-specific behaviour rules
Technical surface
- API surface
- not applicable — file convention
- Backend storage
- file (CLAUDE.md)
- Deployment
- not applicable — not a deployable product
- Embedding model
- not applicable — file convention
- Multi-tenancy
- not applicable — file
- MCP
- not applicable — file-as-memory paradigm (no API surface)
- A2A
- not applicable — file-as-memory paradigm
- OpenTelemetry
- not applicable — file-as-memory paradigm
Compare CLAUDE.md with…
Similar systems
Other file-backed / editor paradigms in the catalog, ranked by inbound references.
- Cline .clinerules T2
.clinerules single file or .clinerules/ directory of numbered .md files (e.g. 01-style.md ). Global at ~/.cline/rules/ . Files version-controlled and AI-editable — Cline can write back. Toggleable per-session via UI.
- AGENTS.md T1
AGENTS.md at repo root (and ~/.codex/AGENTS.md for global); hierarchical concatenation from root to cwd. Persists build/test commands, conventions, architecture overview, security constraints, git workflow.
- Aider CONVENTIONS.md / .aider.conf.yml T2
CONVENTIONS.md (any name, loaded via --read or read: in config); .aider.conf.yml (home / git-root / cwd, last wins); .aiderignore (gitignore-syntax).
- Cline Memory Bank T2
memory-bank/ directory of Markdown: projectBrief.md , productContext.md , activeContext.md , systemPatterns.md , techContext.md , progress.md . Loaded in full at start of every task. Compensates for context-window reset between sessions.
- Continue.dev .continue/rules T2
.continue/rules/*.md at workspace root. YAML frontmatter: name , globs (file-pattern trigger), alwaysApply , description . Loaded lexicographically.
- Cursor Rules T1
.cursorrules + project rules. Semantic chunking + Merkle-tree incremental re-indexing over the codebase. Turbopuffer-backed retrieval at IDE scale.
Related systems
References (1)
- Anthropic Claude (foundation models) depends on at runtime — backend-storage cell: file (CLAUDE.md)