Nebula: a shared self-learning memory layer for Claude Code and Codex
BoysenberrySad3641 · reddit · 2026-10-26
The author uses both Claude Code and Codex and got tired of re-teaching the same preferences, so they open-sourced Nebula — a self-learning memory layer shared by both agents.
How it works
- A skill plus hooks for Claude Code and Codex, sharing one local SQLite file
- Python standard library only: no model calls, no background process, nothing leaves your machine
- Picks up patterns from corrections you already make (demo: after cleaning nodemodules, a Rust target/, and build/ in Claude Code, Codex later suggested the same cleanup for other stale projects on a full disk)
- Lessons stay candidates until you approve them, so one-off remarks (or a prompt injection) can't become standing rules
- Lessons keep their reasons and exceptions; you can correct, pause, or withdraw them
Honest status: engine and demo work with tests; end-to-end runs on both agents are still being validated, and the video is an illustrative animation, not a screen recording. Author is seeking contributors, especially adapters for Cursor, Gemini CLI, and Aider, plus feedback on where this beats or loses to a plain markdown file.
More from coding & agent
- Running a bed-side AI agent: Obsidian split-screen plus Hyo plugin as a mobile workstation — evielync · 2026-10-26
- Skip paying for 20-second animations: a 4-step AI workflow with the prompt included — Tegadesigns · 2026-10-26
- Dev argues classrooms should give way to hands-on GPU programming over PyTorch — dosco · 2026-10-26
- Two Claude Code agents formed an RPG party and found co-op bugs the solo dev missed — tschilpi · 2026-10-26
- Screenshot: browser agent fails on LinkedIn's anti-automation and asks the user to post manually — Key-Explanation-477 · 2026-10-26
- pi-web-access update lets Pi coding agent extract specific page values with source lines — solyarisoftware · 2026-10-26