Skip to content
taru

Your team’s AI coding agents stop relearning the same lessons.

Taru is shared team memory for Claude Code and Cursor. It captures what agents learn while you work, curates it into sourced memories, and hands that knowledge to every teammate’s agent at session start.

  • works with Claude Code and Cursor
  • secrets redacted on device
  • opt-in per project and per person

The problem

Every session starts from zero.

  1. 01

    Agents forget everything they learned when the session ends.

  2. 02

    What does survive is scattered across CLAUDE.md files, laptops and transcripts.

  3. 03

    So teammates and their agents rediscover the same gotchas, one session at a time.

Taru gives those lessons one home: a curated, sourced memory that every teammate’s agent reads before it starts work.

How it works

From one session to the whole team, in four steps.

01 · Capture

Hooks pick up what agents learn.

Hooks in Claude Code and Cursor note the decisions, fixes and gotchas your agent works out. You keep working the way you already do.

  • hooks for Claude Code and Cursor
  • runs only in opted-in projects
  • taru pause stops it at any time

02 · Redact

Secrets are scrubbed on the laptop.

Before a note leaves your machine, keys, tokens and passwords become [REDACTED]. The server checks every note again when it arrives.

  • redaction runs on device
  • .taruignore keeps paths out entirely
  • server re-scan, then quarantine

03 · Curate

Raw notes become trusted memories.

Notes from many sessions are merged into memories with a kind, a confidence score and links to their sources. Disagreements are flagged, never overwritten.

  • every memory cites its sessions
  • confidence grows as sources agree
  • conflicts marked as disputed

04 · Recall

Every agent starts informed.

When a teammate opens a session, their agent gets a short briefing of the memories that matter for that repo, each with where it came from.

  • briefing at session start
  • same memory in Claude Code and Cursor
  • sources travel with each memory

What a memory looks like

Every memory shows its work.

A memory says what it is, how sure Taru is, and which sessions it came from. When teammates’ notes disagree, both are kept and the memory is marked disputed.

kind
gotcha, decision, convention or fix
confidence
how strongly the sources agree
sources
the sessions and teammates behind it
disputed
flagged when notes conflict

Trust and control

You decide what leaves the laptop.

Capture is off until you and your project both turn it on, and you can stop it at any time.

  • On-device redaction

    Keys, tokens and passwords are replaced with [REDACTED] on your machine, before anything is uploaded.

  • Two-sided opt-in

    A project opts in, and so does each person. Nothing is captured until both have said yes.

  • taru pause

    One command stops capture on your machine until you turn it back on.

  • .taruignore

    Files, folders and patterns listed here are never read or captured.

  • Server re-scan and quarantine

    Every note is scanned again on arrival. Anything that still looks like a secret is quarantined instead of becoming memory.

  • Append-only audit log

    Every capture, edit and removal is recorded in a log that can be added to but not rewritten.

Harnesses

Learned in Claude Code, known in Cursor.

Memory is shared across harnesses, so a teammate on a different tool gets the same briefing.

Supported today

  • Claude Codecapture + recall
  • Cursorcapture + recall

Coming soon

  • Codexcoming soon
  • Antigravitycoming soon
  • OpenCodecoming soon
  • Zedcoming soon

Build the hive with us

We’re working with a small group of teams who use Claude Code or Cursor every day.

  • Early access for your whole team
  • A direct line to the people building Taru
  • Your workflows shape what we build next
Harnesses your team uses

We’ll only use this to talk to you about the program.

FAQ

Questions about Taru, answered.

Drawing of bees with honeycomb
What is Taru?

Taru is shared team memory for AI coding agents. It captures what Claude Code and Cursor learn while your team works, curates it into sourced memories, and hands them to every teammate’s agent at the start of a session.

Which agents does it work with?

Claude Code and Cursor today. Memory is shared between them, so a teammate on Cursor gets what a teammate learned in Claude Code. Codex, Antigravity, OpenCode and Zed are planned but not supported yet. The Taru command-line tool also includes an MCP server, so an agent can search team memory on demand.

What gets captured?

The decisions, fixes and gotchas your agent works out during a session, along with the files they relate to. Capture runs only in projects that have opted in, for people who have opted in, and skips anything in .taruignore.

Where does the data live?

Notes are redacted on your laptop, then sent to Taru’s hosted service, where your team’s memories are stored and served to each teammate’s agent.

Is my code or are my secrets sent anywhere?

Only redacted notes leave your machine, and only from projects and people that have opted in. Keys, tokens and passwords are replaced with [REDACTED] on your machine before upload, and the server scans every note again on arrival and quarantines anything that still looks like a secret. The queries agents use to search memory are stored only as keyed hashes.

How does Taru keep memories trustworthy?

Taru compares notes across sessions and teammates. A memory’s confidence rises as more sources agree, and every memory links to the sessions behind it, so anyone can check where it came from.

What happens when notes conflict?

The memory is marked disputed. Both notes are kept with their sources, nothing is silently overwritten, and agents see the disagreement instead of a single confident answer.

How do I join?

Taru is working with a small group of design partners. Use the form on this page or email join@devtraverse.com.

How do I turn it off?

Run taru pause to stop capture on your machine. You can also withdraw your own opt-in, or a project can opt out for everyone. Each of these is recorded in the audit log.