Knowledge Objects: a substrate sketch
Why human–AI collaboration needs a non-linguistic substrate, and what that looks like in practice.
LLMs are powerful and stateless. Humans carry deep context and broadcast it expensively. Every time we sit down together, we re-do the introductions — the project background, the constraints, the things we tried last week and ruled out. The substrate of that collaboration is language, and language is a terrible serialization format for working knowledge.
The problem with conversation as substrate
Chat is great for negotiation but bad for accumulation. A conversation is linear, lossy, and dies the moment the session ends. If you and I worked together for a year over chat, we'd still wake up every morning at zero — re-establishing context that should have been infrastructure.
The usual workaround — "just add it to the system prompt" — doesn't scale. A system prompt is a flat dump. It can't be queried, linked, versioned, or partially loaded. It's a TXT file masquerading as a memory system.
What a substrate would look like
A real substrate would be:
- Structured. Typed units (facts, decisions, patterns, skills, insights), not free text.
- Linkable. Each unit can reference others. The graph carries meaning the units don't.
- Stable across sessions. Stored outside any one conversation, addressable by ID.
- Auditable. A human can mark a unit
verified,disputed, orarchivedwithout rewriting it. - Bidirectional. Both the human and the AI can write to it; both can read from it.
That's the bet behind Knowledge Objects (KOs) — and behind the KO Network I'm using as the working prototype.
What it doesn't try to do
KOs are not a replacement for embeddings, RAG, or vector search. They're not trying to capture every utterance. They're the deliberate layer — the place you put a thought when you've decided it's worth keeping. Most of the value is in the friction of writing it down well.
Where this is going
I'm writing a NeurIPS 2026 position paper around this — arguing that agent autonomy is a property of the environment, not the agent, and that environments need substrates like this to make autonomy legible. The paper is the formal artifact. The lab is where the substrate gets built.
More notes as the work develops.