Korva vs CLAUDE.md
CLAUDE.md is the standard way to give Claude Code project instructions. It is a static Markdown file that gets injected into context on every prompt. That simplicity is also its ceiling — it cannot grow, cannot enforce, and cannot follow your team to another editor.
Where CLAUDE.md stops scaling
- It does not accumulate. Every decision, incident and pattern has to be manually written into the file, or it is simply lost between sessions.
- It costs tokens every prompt. The whole file is injected on each call — relevant or not — so it competes with actual context for the budget.
- It is Claude-only. The moment a teammate opens Cursor or Copilot, the file does nothing for them.
| Capability | Korva | CLAUDE.md |
|---|---|---|
| How knowledge is stored | Live SQLite vault per project + 25 curated Lore scrolls | A single static Markdown file |
| Memory across sessions | Persistent — vault_context loads prior decisions automatically at session start | Starts fresh each session unless manually written into the file |
| Token cost per prompt | Only the relevant observations and scrolls are loaded; vault_compress trims further | The entire file is injected on every prompt |
| Enforcement at commit time | Sentinel pre-commit blocks 10 built-in architecture, security and naming rules | None — instructions are advisory text |
| Multi-editor support | 8 IDEs from one binary, including Claude Code, Cursor, Copilot and Codex | Claude Code only |
| Team knowledge sharing | Encrypted cloud config + portal + Skills Hub + audit log (Teams tier) | Commit the file and hope everyone pulls it |
| Structured workflow | Forge enforces a 5-phase Spec-Driven Development workflow with approval gates | None — the file is plain guidance |
| Setup time | curl + korva init + korva setup — 30 seconds | Create the file — instant |
| License & cost | MIT, free Community tier; Teams from $9/user/mo (launch) | Free, just a file in the repo |
A concrete example: context that does not cost tokens
With CLAUDE.md, a 400-line instruction file is injected on every single prompt — most of it irrelevant to the task at hand. With Korva, vault_context loads only the observations and scrolls that match the open file, and vault_compress can shrink the payload further. The AI gets sharper context for fewer tokens.
Keep CLAUDE.md anyway
Korva reads your existing CLAUDE.md as part of the AI context. Keep it for high-level project description and tone, and let Korva carry the architecture rules, persistent memory and stack knowledge that need to accumulate.
FAQ
Does Korva replace my CLAUDE.md file?
No — Korva reads your existing CLAUDE.md as part of context. Over time you can move accumulating knowledge (decisions, incidents, patterns) into the vault and keep CLAUDE.md for high-level description.
Will Korva reduce my token usage in Claude Code?
Usually yes. Instead of injecting an entire CLAUDE.md on every prompt, vault_context loads only relevant observations, and vault_compress can trim the payload by 20–75%.
Does Korva work outside Claude Code?
Yes. The same vault and rules work across 8 IDEs via the Model Context Protocol — Claude Code, Cursor, Windsurf, Copilot, Codex, OpenCode, Gemini CLI and VS Code.
Replace a static file with a memory system
Install in 30 seconds. Korva works in Claude Code today — and every other editor your team uses tomorrow.