avenoxai/avenoxbeyin

Overview

AvenoxBeyin is a local memory layer for AI agents. It keeps durable information in ordinary Markdown files instead of a chat history: the agent writes decisions, task results and open items into notes, then finds them again in a later session by local search and passes the selected context to whichever supported client is in use. Notes stay readable in Obsidian or any text editor, and the index is kept on the local machine.

The system does not automatically convert an entire chat history into knowledge; what becomes persistent is what you ask the agent to write down. It ships four skills — beyin for finding/saving notes and tasks, beyin-doktor for health checks, beyin-guncelle for updates and rollback, beyin-kontrol for verifying that stored information is still accurate and only runs on request. The shared source lives in .agents/skills, and existing personal skills are not silently overwritten.

Supported clients include Claude Code, Codex, Google Antigravity, Kiro, Hermes Agent, OpenCode, OMP / Oh My Pi and Pi, each with its own hook and trust flow; the README notes that automatic hook delivery on Codex Desktop is not yet considered verified.

Quick start

Prerequisites: Python 3.11+ and a supported AI client (the client and its account are installed separately). Git, pip, a Mem0 account, extra API keys and an always-on server are not required; Obsidian is optional.

1. Pick a folder — a new one or an existing Obsidian vault. Notes do not need to be moved.
2. Open that folder in your AI client so installation runs in the right directory.
3. Paste this message into the agent:

https://avenox.io/beyin.md adresini oku. Bu klasöre ikinci beynimi kur.
Mevcut kurulum varsa notlarımı koruyarak güncelle. Gerekli indirme ve
kurulum adımlarını sen yap. Sonunda örnek bir bilgiyi kaydedip yeni
oturumda geri okuyarak birlikte doğrulayalım.

The agent downloads the stable release package, verifies its checksum and runs the installer; complete the folder/hook trust prompt shown by your client, then open a new session in the same folder (on Codex, review the new hook definitions under /hooks).

If you prefer the terminal, download beyin-v3-X.Y.Z.zip and its .sha256 file from the latest release — not GitHub's "Source code" archive — verify the hash, unzip, and from the unzipped package folder run:

python3 scripts/install_v3.py --vault "/full/path/Beynim"

On Windows use py -3 instead of python3 and a path such as C:\Notlar\Beynim. The installer writes the engine, client connections and the four starter skills into the vault; you can close the package folder afterwards. See the detailed setup guide for per-client instructions.

To confirm it works, ask the agent to save an example decision as a note, then open a new session and ask which choice was made and why, requesting the source note. Optional extras include a local web view (python3 beyin.py web ac, --read-only to view only, bound to this computer and shutting down after 30 minutes idle) and a single-key capture inbox for videos, tweets, articles, mail and Word documents (python3 beyin.py yakala kur), where extracted cards are linked into notes when you ask the agent to process them.

Constraints

• Search is local keyword matching; there is no guarantee that every relevant note is recalled on every question. Important decisions should be explicitly written to notes, and task-result records are the agent's own claims rather than independently verified facts.
• Notes are local Markdown files, but selected context is sent to whatever AI client and model provider you use; "local memory" does not mean the whole AI conversation stays offline. Marking sensitive notes affects Beyin's context and web filters but is not an OS-level access control against an agent you have granted file access to.
• Base local checks do not call a model; work assigned to the agent and added context count toward the client's normal usage. There is no background service scanning memory while clients are closed.
• The capture listener and web view must be reinstalled/restarted after updates, and the vault's working database should not be shared like an ordinary synced file across machines. Migration from V2 and multi-machine use require separate steps.
• Licensed under MIT.

HotRepo's take

The project's only recorded HotRepo snapshot is the 2026-11-13 Python daily list, where it ranked 5th with 46 stars added that day, reaching 436 stars and 74 forks overall. That single snapshot supports the fact of one strong day on the Python list, not a trend; there is no all-language daily rank or longer series to compare against. What stands out technically is the approach — keeping memory in plain Markdown under the user's control and adapting to multiple agent clients instead of locking data into one chat product — and the README's willingness to state limits such as keyword search and unverified Codex Desktop hook delivery, which matters for anyone deciding whether to adopt it.

Project source and documentation