Claude Code Persistent Memory — Never Explain Your Stack Again

Why Claude Code forgets — and why that's expensive

Claude Code operates within a context window. When that window closes — session ends, tab closes, computer restarts — everything is gone. The next session starts cold. No memory of:

• ✗Why you chose this database over that one
• ✗The auth bug you fixed last Thursday and what caused it
• ✗Your team's agreement to always use idempotency keys for Stripe
• ✗The deploy procedure that keeps tripping people up
• ✗What the codebase looked like six months ago and why it changed

The standard workaround is a CLAUDE.mdfile — a markdown document you maintain by hand in your repo. It helps. But it requires constant manual updates, goes stale quickly, doesn't learn from your git history, and only Claude reads it. Cursor and Copilot ignore it entirely.

The fix: cachly MCP server

cachly connects to Claude Code via MCP (Model Context Protocol) — the same standard Anthropic built into Claude Code for external tool integration. Once connected, Claude has access to 89 memory tools that give it permanent, shared, cross-session recall.

Setup: one command, 30 seconds

npx @cachly-dev/mcp-server@latest setup

The wizard auto-detects Claude Code and writes the correct MCP config to ~/.claude.json. One browser click to sign in. No API keys to manage, no credentials in your config:

// ~/.claude.json (auto-written by cachly setup)
{
  "mcpServers": {
    "cachly": {
      "command": "npx",
      "args": ["-y", "@cachly-dev/mcp-server@latest"],
      "env": {
        "CACHLY_JWT": "your-token-here"
      }
    }
  }
}

If you also use Cursor, GitHub Copilot, or Windsurf, the setup wizard detects those too and configures them all in the same pass. Same brain, all editors.

brain_from_git: instant context before session one

After setup, cachly automatically runs brain_from_git — it reads your entire git history and loads the lessons into your Brain:

# What brain_from_git extracts from your git log:
✓ Fix: JWT expiry check before signature validation (commit a3f91c2, 2025-11)
✓ Revert: Redis pub/sub removed — race condition at scale (commit 8b2e441)
✓ Decision: tRPC over REST — type safety across monorepo (commit f9c1203)
✓ Warning: Stripe webhook order not guaranteed — always idempotent

Claude Code arrives at your first session already knowing years of your team's decisions — without you writing a single documentation line.

What a session looks like with cachly

Open Claude Code. On the first tool call, smart_recall fires automatically:

> Good morning. Here's what I remember:
>
> Last session: payments refactor — stripe-webhook.ts
> Open TODOs: idempotency key on charge creation (3 places)
> Team learned: always set idempotency_key for Stripe charges
>   → commit a3f91c2, 2025-11-14
>
> brain_predict: ⚠️ Moderate deploy risk
> The k8s namespace change in PR #441 matches a past
> race condition pattern (incident 2024-08-12).
> Recommend review before deploying to prod.

No prompting. No manual context. Claude just knows.

causal_trace: 30-minute git blame in one call

When Claude Code hits a confusing bug, instead of asking you to explain the history, it can call causal_trace:

// Claude Code calls automatically:
causal_trace({ query: "why does auth fail for expired tokens" })

// Returns in ~2 seconds:
{
  rootCause: "JWT expiry validated after signature — window allows replay",
  fix: "Move expiry check before crypto.verify()",
  fixedIn: "commit a3f91c2 — 2025-11-14",
  command: "git show a3f91c2:src/auth/jwt.ts",
  relatedIncidents: ["incident-2025-09-03"]
}

Thirty minutes of reading git history replaced by a two-second tool call.

cachly vs CLAUDE.md

CLAUDE.md is a good starting point. cachly is what you graduate to when manual maintenance becomes a burden.