Claude Code has no operational memory. Every session starts from zero. You re-explain your stack, paste context, repeat yourself.

Persistia fixes that. On first run, it scans your entire project — code, docs, brand assets, financial models, marketing briefs, configs. Reads your .env files and uses credentials you already have. Stores everything in small, focused files with a navigation index. Then monitors changes via git diff: 1,000 files, 3 changed, it reads 3. Token-efficient by design.

The result: an agent that always knows your full project and stays up-to-date automatically — even as you ship new features, update your positioning, or revise your financial model.

Minimum input. Maximum output.

Open Claude Code in your project root folder, then run:

curl -fsSL https://raw.githubusercontent.com/bernardohcrocha/persistia-for-claude-code/main/setup.sh | bash

Claude scans your project and asks only what it can't find. No forms. No config files. Just a conversation.

Requires: git · Node.js 18+

Optional: GitHub CLI (gh) — automatically creates a private brain repository with cloud backup on every update. Format your machine, git clone the brain repo, continue exactly where you left off.

Scheduler included: installs automatically (launchd on macOS, systemd on Linux).

• Initial indexing — on first run, scans your entire project: code, docs, configs, and environment files. Asks only what it can't find on its own. Uses credentials you already have — no manual reconnection.
• Self-updating memory — runs git diff daily. Only what changed gets read. No unnecessary context loaded.
• Permanent skills — say it once → written to _brain/skills/, loaded at every future session
• Autonomous tasks — schedule any task in plain language, runs automatically with full context. Each task in _brain/tasks/queue.json supports batch_size (cap items per run — prevents timeouts on large lists) and timeout_minutes (default: 5). Easy to customize.
• Proactive scan — every 3 days when idle, it scans metrics, customers, and channels and flags what it notices. Always suggests, never acts unilaterally.
• Live dashboard — open _brain/dashboard.html, auto-refreshes every 5 min
• Shareable memory — its own private git repo. Format your machine, git clone the brain repo, continue exactly where you left off. Shareable with your team.

Unlike Hermes Agent, Agent Zero, or OpenClaw — context updates itself. No manual maintenance required.

Everything in _brain/ is plain text — readable, editable, and version-controlled.

A dashboard shows numbers. This answers questions and executes orders — cross-tools.

• Ask — revenue, churn, at-risk accounts, support patterns — across all your tools at once.
• Give orders — outreach, reports, follow-ups. Executed using credentials you already have.
• Schedule — any task, in plain language, runs automatically with full context.

→ "Which customers dropped usage 30%+ this month? Cross-check their support history and draft a personalized re-engagement message for each."

→ "Every Monday: pull last week's numbers from Stripe, flag anomalies, and queue a follow-up for any account that dropped below quota."

→ "Find signups from the last 30 days with no activity. Visit each company's website and write a personalized email leading with the most relevant pain points."

_brain/
├── .git/             ← isolated brain repository, pushes to private GitHub remote
├── index.md          ← agent reads this first, every session
├── dashboard.html    ← live command center, auto-refreshes in browser
├── core/             ← product, brand, ICP
├── operations/       ← metrics + customers, auto-updated daily
├── skills/           ← permanent rules, created and updated automatically
└── tasks/            ← scheduled tasks queue, managed automatically

→ How the git architecture works

Persistia has two engines.

Engine 1 — every time you open Claude Code

CLAUDE.md is a native Claude Code feature: Claude reads it automatically at the start of every session. Persistia fills that file with a session protocol that runs before Claude responds to anything:

1. Check _brain/inbox/ for new files — process and archive them
2. Run git diff HEAD~1 --name-only — read only what changed since last session, update the brain
3. Read _brain/index.md (the navigation map) — load only the sub-files relevant to the current task
4. Load the relevant skill from _brain/skills/ if the task has one

It doesn't matter what kind of file changed. A new pricing page, a revised ICP doc, an updated financial model, a rewritten onboarding email — Claude reads the diff, updates the brain, and carries that context forward. No manual re-explaining.

Engine 2 — background, even when Claude is closed

Setup installs a system scheduler (launchd on macOS, systemd on Linux) that fires queue-runner.mjs every 15 minutes. When a task is due, the runner calls claude -p "<prompt>" — Claude CLI in headless mode — which opens your project, reads CLAUDE.md, runs the task, writes to _brain/, and auto-commits. No window, no interaction required.

Three tasks run by default: daily brain sync at 9am, weekly review every Monday at 9am, and a proactive opportunity check every 3 days (only when the project is idle — no commits in the last hour, so it never fires while you're actively working).

If the claude CLI isn't found, the runner writes the task to _brain/inbox/ instead — Engine 1 picks it up at your next session.

Your day to day

All schedules are plain text in _brain/tasks/queue.json. To adjust any frequency, just ask — "Run the proactive check weekly instead of every 3 days" — Claude updates it on the spot.

No subscriptions. 100% free. 100% open source.

Claude Code is a product of Anthropic. Persistia is an independent open-source project, not affiliated with or endorsed by Anthropic.

MIT License · 2026