TeamBrain – Git-Native Shared Memory for Claude Code, Cursor and Codex

DonatienMigue1 pts0 comments

TeamBrain — Your team's AI agents share one brain

Copied

teambrain

Docs<br>White paper<br>★ GitHub

Open source · Apache-2.0 · MCP-compatible

Your team's AI agents share one brain.

Claude Code, Cursor, and Codex — reading from and writing to the same team memory. Approved by humans, stored in git, learned from real sessions.

$ npm i -g @teambrain/cli⧉<br>Star on GitHub

5-min quick start →<br>White paper →

memory lifecycle

$ session log

agent discovered: payments retries live in wrapper.ts

title: payments retries

source: session-2026-07-03

confidence: high

Payments retries are implemented in wrapper.ts, not retry.ts.

+ approved

memories/payments-retries-wrapper.md

+ approved

Claude Code

Cursor

Codex

// the-problem.md

You're paying your agents to forget.

Every AI coding session starts from zero. Agents re-read the codebase your teammate's agent mapped this morning; industry cost audits put re-sent context at half or more of agentic token spend. Multiply that across every developer, every tool, every model switch — that's the amnesia tax. Vendors sell you caching to make forgetting cheaper. TeamBrain makes remembering the default.

amnesia tax estimate

Team size: 20

Monthly AI spend per dev: $150

≈ $1,650/month

spent re-reading what your team already knows — estimate, based on published context-reuse audits.

// how-it-works.md

How it works

01

Capture.

Lightweight hooks record what sessions did — metadata only, redacted on-device, auditable with one command.

02

Distill.

A CI job turns repeated struggles and discoveries into proposed memories — as pull requests.

03

Serve.

Approved memories reach every agent, in every tool, over MCP. New hire? Day-one context.

teambrain/proposals-2026-07: 3 new memories

Copy ⧉

+ memories/payments-retries-wrapper.md

+ memories/auth-session-refresh.md

+ memories/flaky-test-quarantine.md

3 files changed

// trust.md

Governed like code, because it is code.

Every memory is a markdown file in your repo, approved through your review process. No TeamBrain servers — your git host is the transport, your CI is the compute, your LLM key does the distilling. No raw prompts recorded, ever. No individual metrics, ever. git clone is your exit strategy.

No TeamBrain servers — your git host is the transport, your CI is the compute, your LLM key does the distilling.

No raw prompts recorded, ever.

No individual metrics, ever.

git clone is your exit strategy.

Why teams, why now

70% of engineers already run 2–4 AI tools. Per-vendor rules files drift; session learnings evaporate; leads can't see what works. TeamBrain is the neutral layer none of the vendors can be.

// codemap.md

Stop paying your agents to re-read the repo.

v1.1

Agents waste half their tokens re-exploring code your teammate's agent already mapped. CodeMap builds a shared, always-current map of your codebase in CI — updated incrementally on every merge — so an agent asks "where do webhook retries live?" and gets a 200-token answer instead of a 40,000-token exploration. Same repo, same MCP tools, no extra work.

~40k tokens exploring

read src/webhooks/index.ts

read src/webhooks/queue.ts

read src/lib/http/client.ts

read src/lib/http/retry.ts

read src/payments/wrapper.ts

read src/payments/hooks.ts

read src/config/queues.ts

read src/workers/dispatch.ts

read src/workers/retry-policy.ts

grep "retry" src/**

~200 tokens, from CodeMap

webhook retries → src/workers/retry-policy.ts, dispatched from dispatch.ts

Open source, Apache-2.0.

Five minutes to your first shared memory.

Quick start<br>GitHub<br>White paper

GitHub<br>Docs<br>White paper<br>Apache-2.0

Team cloud & FlightDeck practice insights — coming. Aggregate-only, by design.

read teambrain payments retries memories code

Related Articles