One brain. Many agents. Shared memory.
Puff gives a team of real coding agents one shared graph memory. Every lesson is tied to the code it came from, so the next agent starts with the answer instead of repeating the mistake.
Claude Code, Codex and OpenCode read and write one shared graph memory.
Tasks
2 running · 1 queued · 1 verifiedliveWhat you can do in Puff
Main tools are on top. Click one to see what it does. More tools are listed below.
Memory tied to the exact symbol
Every failure and decision points at a real function or class, so memory can be found, checked, and invalidated when that code changes.
- Files, functions and calls parsed with tree-sitter
- Failures linked to the symbol they happened in
- Stale lessons get flagged when code changes
I pin every lesson to the exact symbol it came from.
[Code Node]
src/auth/login.ts
symbol · login()
[Task]
Add rate limiting
status · planned
[Lesson]
Redact credentials in audit log
past fix · verified
Each edge carries a reason · graph traversed, not grepped
// problem → puff matrix
Amnesia vs. shared graph memory
Rediscover the codebase every time
Targeted ContextPack compiled from the tree-sitter AST graph
85% token reduction
Repeat past test failures
Failures linked to the exact Symbol (function / class) in FalkorDB
0 repeated traps
Vendor lock-in & amnesia
Claude Code, Codex and OpenCode read and write to one graph
100% shared memory
Nobody routes work by strength
Routing by declared strengths, adjusted by past successes
Explainable picks
Play with Puff's Context Engine
Play with Puff's Context Engine
src/auth/login.ts:12-48
- routes/auth.ts::postLogin
- session.ts::Session.start
"test: password appears in audit log" → fixed by: redact credentials before audit()
// Auto-generated by Puff Context Compiler
TASK: Add rate limiting to login endpoint
TARGET CODE: src/auth/login.ts:12-48
PAST LESSON: "test: password appears in audit log" → fixed by: redact credentials before audit()
ACTIVE DECISION: Tokens are stored in httpOnly cookies, never localStorage
Get Puff for your team
Get Puff for your team
Run Claude Code, Codex and OpenCode on a single repository with shared graph memory.
New in this release: a live Tasks view with the graph's reason for each task and one-click “Open in” your agent's terminal, a Memory screen with every lesson's trust, and fixes for projects whose checks already fail. Windows 10 or 11, 64-bit, about 115 MB. Already installed? Run the new installer and it updates in place, keeping your projects and memory. The installer is not code-signed yet, so Windows may warn about an unknown publisher: choose More info, then Run anyway. macOS and Linux: run from source with the command above. Release notes