Building a multi-agent shared knowledge base with two GitHub repos — and one memory pitfall
gregbarbosa · x · 2026-10-18
A developer shares how his team built a centralized knowledge base that all agents can read, to avoid lock-in to one platform:
- Two private GitHub repos of plain markdown: a "company brain" (brand, customers, products, sales, ops, finance) and a "personal brain" (health, investments, family).
- A single AGENTS.md every agent reads first.
- A daily "knowledge base audit" skill that diffs the day's work against both brains and suggests edits.
A replier flags a real pitfall: memories cross-trigger safeguards between agents. His GLM found a Roku TV exploit, and Claude pulled that memory while debugging apps, started safeguarding follow-up requests, and eventually refused to continue.
More from coding & agent
- Model-agnostic harnesses will never get the most out of a model — gethackteam · 2026-10-18
- Google open-sources Harness²: recursive agent harnessing lifts GLM-5 and Gemini 3.5 Flash by up to 11.8 points — dr_cintas · 2026-10-18
- exe.dev ships API/CLI explorer for its 1-second VM platform — davidcrawshaw · 2026-10-18
- Using Codex to Reorganize a Google TV: Coding Agents in Unexpected Places — brandon_galang · 2026-10-18
- Monorepos went from 'suck' to 'awesome' because AI can navigate them — gethackteam · 2026-10-18
- Non-programmer on coding agents: 'I've never written code, I just trust the process' — flowersslop · 2026-10-18