@daksh_jaitly: I built mex because coding agents kept rereading the same repo instead of preserving what they learned. 1,200 GitHub st…

I built mex because coding agents kept rereading the same repo instead of preserving what they learned.

1,200 GitHub stars later, v0.7.0 adds a local code graph. In our repo benchmark, it returned 90.7% less context than grep with 100% expected-symbol recall. (Thread)

(2) The idea behind mex is simple:

A living wiki for your codebase, maintained by coding agents.

Agents write architecture, decisions, conventions and reusable patterns into repo-local Markdown, so future sessions inherit what previous agents learned instead of starting from zero.

(3) v0.7.0 puts a deterministic code graph underneath that wiki.

Tree-sitter extracts symbols and relationships into a local SQLite database.

No embeddings. No AI required to build it.

Supports TS/TSX, JS/JSX, Python and Rust. (and many more coming)

here is an example on the hono codebase

(4) For a task, the agent can run:

mex graph scope “trace the authentication flow”

Instead of dumping full files, mex returns a compact neighbourhood of relevant symbols, callers, callees and relationships under a hard token budget.

Source is expanded only when needed.

(5) On the mex repository benchmark:

• 10.74× smaller than grep top-3 • 90.7% less returned context • 100% expected-symbol recall • 5/5 real-agent tasks completed • 0/5 needed fallback Read/Grep with minimal graph context

(6) The weirdest result:

Giving the agent less source upfront worked better. Minimal graph context + targeted graph get calls needed no Read/Grep fallback.

Inline-source mode needed fallback on 4/5 tasks. A map first. Source only when needed.

(7) The graph also connects the wiki back to the code. A Markdown claim can point to the exact function that makes it true. If that function changes, moves or disappears, mex can flag the affected knowledge as stale.

Small benchmark, public methodology, fully open source:

Checkout the repo: https://github.com/mex-memory/mex

And also join our discord (we’re planning some fun stuff here) https://discord.gg/FEdNsQ4Qt4

90% less tokens.