Skip to content

Open source · local-first · agent-native

Memory for your coding agents, across every tool you use.

arca turns your Claude Code, Codex, and Cursor sessions into searchable memory — one command away, in any agent.

curl -fsSL https://tryarca.ai/install | bash
View source

How it works

Three commands between you and total recall.

01

Index

$ arca index

arca watches your local Claude Code, Codex, and Cursor stores and normalizes every session into anchored sections — stable, citable excerpts with real message anchors. Incremental, append-aware, and fast.

02

Retrieve

$ arca context "oauth regression"

Lexical recall from SQLite FTS, semantic recall from a warm daemon, and a mandatory reranking pass. You get the three sections that matter, not a transcript dump. Typical query: under 50ms.

03

Continue

$ arca init all

One command installs native skills for all three agents. From then on, your agents check prior work before broad exploration — automatically, in whatever tool you happen to open.

What you get

Memory infrastructure that respects how you actually work.

Anchored sections, not dumps

Retrieval returns citable excerpts — the exact exchange that matters, with a stable anchor you can open later. Full transcript export exists, but only when you ask for it.

── codex:019d…#u5-a5 · 6d ago ─────────────
user: should we rotate refresh tokens on every request?
agent: no — rotation per request breaks concurrent tabs. we keep a 15m window instead, revoke on reuse. ← the decision

Local-first by design

Your sessions never leave your machine. The index, corpus, and daemon all live under one directory. Team sync is opt-in and scoped to sections you explicitly share.

~/.arca
├── index.sqlite # metadata + fts
├── corpus/ # anchored sections
├── qmd-shards/ # semantic index
└── arcad.sock # warm daemon

Strict filters

Hard filters run in SQLite before any semantic ranking — so “this repo, last week, no subagents” means exactly that.

--provider codex--since 7d--repo-scope strict

A warm daemon

Semantic search and reranking stay resident, so nothing cold-starts. New sections are searchable the instant they’re indexed.

$ arca daemon status
● arcad ready · semantic warm · 12ms

Built for agents first

Skills for Claude Code, Codex, and Cursor. An MCP server for everything else. Plain-text output any model can read.

$ arca init all
✓ claude · codex · cursor · mcp

Things your agents can now answer

“why did we pick pnpm over bun here”“what broke the last time we touched auth”“how did sam fix the flaky e2e suite”“what was tested before the billing migration”“the session where we profiled the slow query”“which approaches we already rejected for caching”“what the agent found in yesterday's debugging run”“the reasoning behind the schema denormalization”

For teams

Your teammate’s 2am debugging session, available at your 9am standup.

Solo, arca is perfect recall for one machine. On a team, it becomes shared engineering memory: every agent session anyone runs — the decisions, the tests, the dead ends — searchable by everyone else’s agents. New engineers inherit context instead of archaeology.

See team pricing

FAQ

Questions, answered.

Everything else lives in the README.

Does my session data leave my machine?+

No. arca is local-first: indexing, search, embeddings, and reranking all run on your machine, under ~/.arca. The paid team tier adds opt-in sync of anchored sections you explicitly choose to share — never raw provider stores by default.

Which agents are supported?+

Claude Code, Codex, and Cursor are first-class: arca parses their native local session stores and ships a skill for each. Any other MCP-capable agent can use arca through the MCP server.

Is it really open source?+

Yes. The CLI, daemon, indexing pipeline, and agent integrations are open source on GitHub. You can run everything forever without paying. Paid plans cover team sync, hosted memory, and support.

What does retrieval actually return?+

Anchored sections — compact, citable excerpts with stable message anchors (like codex:019d…#u5-a5), ranked by lexical recall, semantic recall, and a mandatory reranking pass. Full transcript export is one flag away when you need it.

Will it slow my machine down?+

Indexing is incremental and append-aware: only changed session tails are reparsed. A single warm daemon handles semantic search so queries don't cold-start. Typical retrieval is under 50ms.

How does team memory handle privacy?+

Sharing is explicit and section-scoped. You (or a policy) choose which anchored sections sync to the team. Everything else stays local. Attribution is preserved so you know whose reasoning you're reading.