Getting Started
Your AI forgets everything between sessions. TRW gives it persistent memory, structured workflows, and quality gates. Install once — every session builds on the last.
Without TRW vs. with TRW
- Learnings lost every session
- No checkpoints, rework from scratch
- Skip tests, ship bugs
- No quality gates, defects cascade
- Context evaporates on compaction
- Knowledge compounds across sessions
- Checkpoints survive interruption
- Automated validation gates
- Quality audit before delivery
- Discoveries persist permanently
Three steps to your first session
Run the installer
One command. Auto-detects Claude Code, OpenCode, or Cursor.
Start a session
Run trw_session_start(). Every session after this one starts smarter.
curl -fsSL https://trwframework.com/install.sh | bashWorks with Claude Code, OpenCode, and Cursor. Full walkthrough
What gets added
.mcp.jsonMCP server config for your AI coding tool
CLAUDE.mdProject instructions — high-impact learnings auto-promote here
.trw/Learnings, analytics, session state — the memory layer
Zero infrastructure. No database, no cloud service, no server. Everything runs from flat files in your repo.
What's included
What TRW does
Five capabilities. Click any to see the code.
Persistent memory that compounds
Session 50 surfaces what session 3 discovered — automatically.
trw_learn(summary="SQLite WAL mode required for concurrent reads")
# 20 sessions later, recalled automatically
trw_session_start()
# => Loaded 47 learnings (3 high-impact for this context)Structured workflows
Typo fixes flow in seconds. Security migrations get full review.
trw_init(task="fix typo in README")
# => Ceremony: MINIMAL (Implement → Deliver)
trw_init(task="add OAuth2 to API")
# => Ceremony: FULL (all 6 phases)Agent teams with zero merge conflicts
File ownership resolved before the first line is written.
TaskCreate(
assignee="impl-auth",
files=["src/auth.py", "tests/test_auth.py"],
playbook="Implement OAuth2 per PRD-CORE-055"
)Build verification before delivery
pytest + mypy + coverage in one command. Fails block delivery.
trw_build_check(scope="full")
# => 4,127 tests passed, mypy clean, coverage 91%
trw_deliver()
# => Learnings persisted, run closed, CLAUDE.md syncedRecovery from context compaction
Checkpoints save to disk. New sessions resume, not restart.
trw_checkpoint(message="Auth routes implemented, tests passing")
# Context compacts... new session starts
trw_session_start()
# => Recovered: "Auth routes implemented, tests passing"Built by agents, verified by evidence
TRW governs its own development. Every number below is real output from agents using TRW to build TRW.
Source-available under BSL 1.1 — converts to Apache 2.0 after 4 years.