Why credit matters
Almost nothing in our agent operating system is a new idea. It's a combination of patterns other people worked out and gave away. They deserve the credit, and you deserve the originals, which explain things better than a summary can.
We're not affiliated with any of these projects, and naming them isn't an endorsement in either direction. Where we describe a project, the description is our reading of its public README at the time of writing. Check the repository for the current state.
The projects
| Project | What it is | What we took from it | License |
|---|---|---|---|
| GBrain (Garry Tan) | A searchable, linked memory for agents and teams, served to coding agents over MCP | Markdown as the source of truth with a rebuildable index, entity pages, brain-first lookup, scheduled maintenance, cited answers, human approval for schema changes | MIT |
| LLM Wiki (Andrej Karpathy) | An idea file for an LLM-maintained wiki between you and your raw sources | The three layers (immutable sources, agent-maintained pages, a schema file), ingest, query and lint, an append-only log, filing good answers back | Public gist |
| autoresearch (Andrej Karpathy) | Agents running unattended training experiments overnight | Bounded loops: one editable surface, a fixed budget per run, a single score, keep or discard | MIT |
| ai-memory | A self-hosted shared memory and handoff server for coding agents | Immutable raw captures, compiling observations into pages, attributed and audited writes, page restore, typed handoffs with leases | MIT |
| Basic Memory | A local-first knowledge base that people and agents edit together | Notes as entities with observations and typed relations, guarded writes, explicit edits instead of overwrites | AGPL-3.0 |
| COG Second Brain | A git-backed second brain run by skills and worker agents | Read-only verifiers, no durable writes without approval, last-verified dates and confidence, daily, weekly and monthly cadences | MIT |
| Knowledge Work Plugins (Anthropic) | Role-based plugins bundling skills, commands and connectors | Packaging procedures by role in plain, editable files | Apache-2.0 |
| Obsidian Skills | Skills that teach agents Obsidian's markdown formats | Plain markdown that people own, with skills to work on it | MIT |
| OpenAI Skills | An early catalog of agent skills, now marked deprecated | Curation tiers for skills, and a reminder that tools get retired | Per skill |
| Khoj | A self-hostable assistant that searches your documents and runs scheduled research | Scheduled research and the case for keeping your files on your own servers | AGPL-3.0 |
The standards underneath
- Model Context Protocol: one way for every agent to reach the same tools and memory, so knowledge doesn't get stuck inside one product.
- Agent Skills: an open format for packaging procedures that several agent tools can load.
Where we made different choices
Borrowing well means knowing where to change things for a company that serves customers in regulated industries:
- Human approval on durable writes is the default, not an opt-in. Some projects auto-approve. We don't.
- Permissions stay on. The research setup in autoresearch suggests disabling them. For agents touching customers and money, we keep every limit in place.
- One brain per portfolio company. Playbooks are shared. Customer data isn't.
- Every claim has a source and a last-verified date, because our pages get quoted to buyers.
A note on licenses
Licenses differ. MIT and Apache-2.0 are permissive. AGPL-3.0 has obligations if you modify the software and offer it over a network. If you plan to use any of these projects' code, not just their ideas, read the license first.
Read the originals. Then build your own, and give something back.
Building in a legacy industry?
We bring the agent crew and the distribution. You bring what you know about the industry.