digitalOS (Personal Brain OS)
ActiveA Git-based personal operating system for AI agents: 80+ plain-text files, no database, ~87% fewer tokens per task through progressive disclosure.
Overview
digitalOS is a personal operating system for AI agents that lives inside a Git repository. Clone it, open it in Cursor or Claude Code, and the agent has the context it needs: voice, goals, contacts, content pipeline, research, and a running log of decisions and failures. No database, no vector store, no API keys, no build step. Just files an agent reads natively.
It is the system behind the essay “The File System Is the New Database”. The public artifact is the digital-brain-skill example in the Agent Skills for Context Engineering repo.
The decision: no database
The counterintuitive choice was using the file system as the store. Three formats, each for a reason:
- JSONL for append-only logs (posts, contacts, interactions, ideas, metrics, experiences, decisions, failures). Every file starts with a schema line so the agent knows the structure before reading the data.
- YAML for configuration (goals, values, learning, circles, rhythms, heuristics).
- Markdown for narrative context (voice guides, research, templates, drafts).
Roughly 80+ files total: ~11 JSONL logs, ~6 YAML configs, 50+ Markdown documents. Because it is Git, every change is versioned and traceable, and nothing is lost.
How it stays cheap
The architecture is progressive disclosure with module isolation:
- Three-level loading:
SKILL.mdtoMODULE.mdto the data files. The agent loads only the module relevant to the current task instead of the whole repo. - Module separation: identity, content, knowledge, network, operations, and automation are independent. One domain’s data is not noise for another.
- Append-only memory: logs are never rewritten, so patterns across decisions and failures stay inspectable.
The result on a representative content task: about 650 tokens of loaded context versus roughly 5,000 without optimization, around an 87% reduction. The point is not the absolute number; it is that the system scales by adding files without inflating what enters the context window.
How it was built
The system was designed by applying the context engineering skills from the Agent Skills repo: context fundamentals for progressive disclosure, context optimization for the token budget, memory systems for the data layer, tool design for the automation scripts, and project-development for the staged, idempotent structure. The traceability from each design decision back to a specific skill is documented in HOW-SKILLS-BUILT-THIS.md.