Different agents.
Shared context.

Useful context should not stop at the end of a conversation.

Working with several AI tools creates a practical coordination problem: notes, decisions and corrections can stay attached to individual sessions instead of informing the next piece of work.

I connected existing agent tools to a shared, Git-versioned body of Obsidian notes. The goal was continuity: retrieve relevant material, delegate a bounded task and preserve what is worth reusing.

1. Make the knowledge reusable, not tool-specific

   Git-versioned Obsidian notes provide shared context for Codex, Claude Code and Hermes. Versioning makes changes reviewable rather than treating the knowledge base as an invisible, ever-changing memory.

2. Use more than one route to retrieval

   Keyword search handles exact terms. Semantic retrieval finds related meaning. Graph links connect concepts across notes. Together they make the knowledge accessible in more than one way.

3. Delegate tasks without delegating accountability

   Specialist agents support focused work. Human approval, monitoring, backup and recovery remain part of the operating approach; automation does not remove the need to inspect important outputs.

4. Turn corrections into better working habits

   A feedback workflow captures useful corrections and updates reusable skills and workflows. This is process improvement, not model training.

The work is in how the pieces fit together.

My contribution is system design and integration: shaping how existing tools share context, route work and incorporate feedback. The underlying models and frameworks remain the work of their respective creators.

Experimental graph-memory work stays separate from production memory while I explore retrieval quality and source provenance.

This page shares the design approach, not the implementation. Source code, prompts, configurations, knowledge contents and deployment details stay private.

Capability is only useful with a way to check it.

The most interesting part is not simply adding another tool. It is deciding what should be remembered, what should be retrieved, what can be delegated and what still needs a person’s judgment.

That is the kind of applied AI work I want to keep doing: connecting technical capabilities to a concrete way of working, with clear limits and a route to recovery.