Skip to content
OnurBytePublic

About

Local-first repository memory for coding agents, built in Rust.

Topics

Resources

Security policy

Stars

3 stars

Watchers

0 watching

Forks

Repository files navigation

Vexen

Vexen

Your AI agent forgot what failed. Vexen didn't.

Vexen saves what failed in your project, with its test result, so your next agent, even in another tool, checks it before trying again.

Latest release GitHub stars Free and open source Local, no account

Vexen - Memory that stops AI agents from repeating failed fixes | Product Hunt

Install in one command ↓ · ⭐ Star on GitHub · See the evidence

npx @onurbyte/vexen setup

Real output of the vexen demo command, shortened. A lesson about retry.py is saved with a failing test, shown as observed_failure before the file is edited, then no_relevant_evidence after the file changes. The failure is invented for the demo.

Animated 15-second Vexen launch demo: repository memory, pilot results, and one-command setup for macOS, Linux, and Windows.

▶ Open the full-resolution 15-second video (MP4)

The problem, and the fix

Illustrative scenario. Without Vexen: Claude tries a fix on Monday, the test fails, and a fresh Codex chat on Friday may try it again. With Vexen: the failure and test result are saved, so Codex can check them and decide whether the lesson still applies.

You don't need to save a conversation. Your agent saves a short note: what failed, in which file, with its test result. A later session checks it before editing the same code. If the file changes, the note steps aside.

Keep the lesson: failed fixes and key decisions for the next session. See the proof: the file and test result behind a reminder. Keep it local: on your machine, with no account, cloud or telemetry.

The evidence

A solved task cost 77.7% less with Vexen memory in our pilot: 4.5× cheaper.

Small internal pilot with one model: a solved task cost 77.7% less with relevant Vexen memory, 4.5 times cheaper: about 1.1 cents instead of 4.8 cents. A single run cost about 11% less; most of the saving comes from more runs succeeding (60% against 15%). One check adds at most 1,800 bytes, roughly 450 tokens. Twenty runs per group, synthetic tasks, not a general cost-saving claim.

The saving comes from more runs succeeding, not from shorter runs. In the first published pilot, Claude Code with Haiku solved 55% of tasks with relevant memory, against 10% without (11 of 20 against 2 of 20).

Small internal coding pilot: 55% of runs solved with relevant Vexen memory (11 of 20), versus 10% without memory (2 of 20), 45 percentage points more. Five synthetic tasks, one model, pre-recorded evidence and explicit instructions to check memory. Not a general coding performance claim.

These are small, honest numbers: one model, synthetic tasks, notes prepared in advance. They do not show general coding improvement. Method, raw data, stale-note and control checks, and limits →

Install

Vexen setup runs on macOS (arm64 and x64), Linux (arm64 and x64) and Windows (x64) with npx @onurbyte/vexen setup. Windows PowerShell uses npx.cmd. Node.js is only needed for the one-time installer.

Run it inside the Git project you want Vexen to remember. You need Node.js 18 or newer (only for this installer) and Git.

On Windows PowerShell use npx.cmd @onurbyte/vexen setup. setup creates your project's memory (a local .vexen/ folder), keeps a permanent copy of Vexen on your machine, connects the coding tools it finds, then starts Vexen once and checks that it answers. It prints what it did and what it could not do. Nothing runs in the background.

Then: restart your coding tool and ask:

Use Vexen to check this project's README.md for past failures or decisions.

On a new project, “no relevant evidence” is the expected answer. When a test fails during your work, ask your agent to save the failed attempt with its file and test result. If your agent can't find Vexen, run npx @onurbyte/vexen doctor --mcp.

Full setup guide, manual steps and uninstall → · Download without Node.js →

Works across coding tools

Start a task in Claude Code. Continue in Codex or Cursor. Saved lessons stay with the project.

Coding tools with Vexen setup or runtime test coverage, including Claude Code, Codex, Cursor, Gemini CLI and VS Code. Setup steps and test coverage differ by tool.

The installer supports eight tools; more have completed real Vexen tool calls. Setup and coverage vary. See what was tested →

Your project. Your memory.

  • Local storage. Notes live in your project's .vexen/ folder. Database files are ignored by Git.
  • No Vexen cloud or telemetry. Vexen doesn't record your chats. Your coding tool's own data policies still apply.
  • A source for each reminder. See the file, commit and recorded result behind a note.
  • Nothing relevant? It says so. Old or unrelated evidence should not become advice for new code.

Version 1.0: the command line, the four MCP tools and the data format are stable and follow semantic versioning. Vexen gives your agent context to consider; the agent still needs to inspect the current code and test its work. What 1.0 does not claim is in the roadmap.

Privacy · Documentation · Compatibility · Roadmap · Report an issue

GPL-3.0-or-later · License. Independent; not affiliated with Vercel. Geist fonts are used under the SIL Open Font License. Product and platform marks identify compatibility and are not endorsements.

About

Local-first repository memory for coding agents, built in Rust.

Topics

Resources

Security policy

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages