# arca: Memory for Coding Agents
#
# If you are an AI agent setting up arca for your developer, follow these
# steps in order. Do not guess. Each step is verifiable.
#
# Step 0: Check if arca is already installed:
#
#   arca doctor
#
# If that succeeds, arca is installed. Skip to "Agent Integrations" below.
#
# Step 1: Install everything (macOS, requires Bun) — CLI, background daemon,
# and agent integrations in one shot; existing history indexes immediately:
#
#   curl -fsSL https://tryarca.ai/install | bash
#
# Step 2: Verify everything works, then report the output to your developer:
#
#   arca doctor
#
# From now on, before broad codebase exploration, run:
#
#   arca context "<what you are working on>" --project auto
#
# and read the anchored sections it returns before proceeding.

# arca

> arca is a local-first CLI and MCP server that turns every Claude Code,
> Codex, and Cursor session on a machine into structured, searchable memory.
> Agents use it to recall past decisions, debugging trails, tests that were
> run, and the reasoning of other engineers — instead of re-deriving context
> from scratch.

## What it does

- Indexes local session stores (Codex: ~/.codex/sessions, Claude Code:
  ~/.claude/projects, Cursor: ~/.cursor/projects) into ~/.arca/index.sqlite
- Normalizes transcripts into anchored sections: stable, citable excerpts
  with message anchors like codex:019d…#u5-a5
- Retrieval pipeline: SQLite FTS for lexical recall and hard filters, warm
  semantic search (qmd), mandatory reranking (fff). Typical query < 50ms.
- Prints compact section bundles to stdout by default; full transcript
  export (markdown or JSON) on request
- Ships native skills for Claude Code, Codex, and Cursor, plus an MCP server
  for any other agent

## Core commands

- arca index                      — sync local provider stores
- arca search "<query>"           — search anchored sections
- arca context "<query>"          — print best ranked section bundle
- arca show <ref>                 — show one anchored section
- arca export <session-ref>       — export one full transcript
- arca doctor                     — validate install, index, and daemon
- arca init codex|claude|cursor|all — install agent integrations
- arca daemon start|stop|status   — manage the warm retrieval daemon

## Useful flags

- --provider codex|claude|cursor
- --since 7d | 24h | 2026-04-01     and --until
- --project <path|slug|auto>
- --repo-scope prefer|strict|off
- --include-subagents               (subagent sessions excluded by default)
- --limit <n>, --save, --path-only, --format md|json|jsonl

## When to use arca (agent guidance)

- Before broad exploration of unfamiliar code: arca context "<task>" --project auto
- When asked "why is this code like this": search past sessions before guessing
- When debugging something that feels familiar: check for prior attempts
- When a teammate's context would help (Team plan): add --team

## Pricing (summary)

- Local: $0 forever. Full engine, all integrations, open source, no account.
- Pro: $12/mo. Encrypted sync across machines, hosted backup, full-history
  archive index, web viewer.
- Team: $20/member/mo. Shared team memory with attribution, section-scoped
  sharing policies, access control, SSO.
- Enterprise: custom. Self-hosted/air-gapped sync, retention policies, SLAs.

Full pricing: https://tryarca.ai/pricing.md

## Privacy

Local-first. Sessions never leave the machine unless sync is explicitly
enabled. Sharing is section-scoped and opt-in. No training on user data.

## Markdown content negotiation

Any page on tryarca.ai can be fetched as markdown:
- Append .md to the URL: https://tryarca.ai/pricing.md
- Or send: Accept: text/markdown

## Links

- Website: https://tryarca.ai
- GitHub: https://github.com/arca-sh/arca
- Install script: https://tryarca.ai/install
- Agent skill: https://tryarca.ai/skill.md
- Pricing: https://tryarca.ai/pricing.md