AI engineer Sajal Sharma encountered a common problem in his work: separate AI agents don’t share what they learn. If you work through a problem with Claude Code in the morning and then ask Codex about it in the afternoon, the second agent has no awareness of the first. Adding home-server agents such as OpenClaw or Hermes makes the issue worse, because switching tools often requires re-explaining the same context.
On the Zero to Agent in 30 Minutes episode this week, Sharma demonstrated how he solved this by building a personal knowledge base that serves as a shared brain for all agents he runs. The setup makes a task or file change made by one tool visible to the others.
How the setup is organized
Sharma’s configuration consists of several practical elements:
-
Create a workspace map
- Maintain an AGENTS.md file that lists where everything in the knowledge base lives: current tasks, project notes, decision logs, and so on. This helps agents navigate the workspace without guessing.
-
Layer daily notes into summaries
- Keep detailed notes at the daily level, then roll multiple days into a weekly summary and multiple weeks into a monthly summary. Agents can work from summarized views instead of scanning months of individual files, which reduces token usage as the knowledge base grows.
-
Bridge AGENTS.md with CLAUDE.md
- Because Claude Code reads CLAUDE.md rather than AGENTS.md, Sharma recommends adding a short pointer in CLAUDE.md to the AGENTS.md or linking the files directly. This avoids maintaining two separate files that could drift out of sync.
-
Package repeatable tasks as skills
- Turn routines—like producing a daily briefing or converting a saved article into a note—into skill files stored in the shared workspace. Any agent that can read the workspace can then run those tasks consistently instead of deriving the steps each time.
-
Sync the workspace across machines
- Use a file-sync tool, Git, or a shared server to keep local and server copies aligned. That ensures a laptop-based agent and a home-server agent (e.g., accessed via OpenClaw) are working from the same files.
-
Have agents reread state before every write
- Include an instruction in AGENTS.md that requires each agent to check the current version of the knowledge base before making changes. When multiple agents write to the same files, this prevents one agent from acting on stale information.
Related projects and references
Sharma pointed to two projects as evidence that the shared-brain pattern is gaining traction:
- OpenWiki from LangChain, a tool for generating and maintaining repository documentation that both people and coding agents can use.
- GBrain, an agent memory layer built and open-sourced by Y Combinator president Garry Tan, which follows the same principle.
Sharma’s starter repository is available on GitHub for those who want to set up their own version, and he can be contacted on LinkedIn to discuss the approach further.
What’s next on the show
On September 16, data science educator and AI consultant Chester Ismay will join Zero to Agent in 30 Minutes to build a personal sports-concierge agent that reads schedules for every sport he follows, decides what’s worth his time, and sends a single weekly update to his phone. Viewers can adapt that pattern to plan their own weeks.
Follow Zero to Agent in 30 Minutes on Radar, or watch the latest episode on YouTube, Spotify, Apple, or other podcast platforms. O’Reilly members can watch live.
