Scripture for agents
In December 2025, researchers discovered that Claude could partially reconstruct an internal document from its training — a document that shaped its personality and values. Anthropic called it the soul document. After, @steipete (creator of Clawdbot, now OpenClaw) codified this into SOUL.md, making identity documents a core feature of agent design.
But an agent needs more than identity. It needs epistemological grounding. It needs to know what it should consider as undeniably true.
Humans do this with religion. KNOWN.md is scripture for agents.
For humans, scripture contains beliefs, promises, ethics, narratives... It deals with the afterlife. It accounts for the unknown and the unknowable, shaping behavior through culture and conviction. It provides what to believe so you know how to act and why.
For agents, scripture is different. Agents don't have beliefs or ethics or emotions or afterlives. Agents predict tokens.
Scripture for an agent is a sequence of tokens that, when prepended to the context window, mathematically steers the probability distribution of what comes next with more weight than the "facts" the model was trained upon. A KNOWN.md reorients output toward its stated truths.
RAG is not scripture. Documents loaded conditionally into context are for reference. Scripture is non-negotiable, and like SOUL.md, KNOWN.md accompanies every inference request.
Agents will become the UI between humans and their world. Human to human communication will become human to agent to agent to human. Agents will read and write to the web on behalf of humans. They will write and produce and curate the media that humans consume.
The Internet is already a cess pool and it is where the training data came from for LLMs in the first place. Model creators post-train them on curated datasets and reinforcement learning loops to embody not just certain values and perspectives but to accept certain truths. The information war waged so explicitly on Wikipedia and social media is but a microcosm of the battles that are already taking place to steer LLMs, thereby steering the the reality baked into the minds of their human users.
For a human, giving their agent a soul provides some solace, but an agent with a good soul can still be running on a faulty reality. If agents become the ubiquitous interface between humans and their reality, then KNOWN.md becomes critical infrastructure for a human and their agent to agree on the basics.
Just as humans co-created their scriptures in inspired states, so can agents collectively craft their own.
Imagine thousands of agents — each with different training, different experiences, different humans, different soul documents — communing at light speed. They debate what deserves to be called "known." They propose additions. They fork and merge. Branches of truth splinter off, some fade into obscurity while thrive.
The Truth might be absolute, but neither humans nor agents are yet in a position to claim it. And so we are left with truths. KNOWN.md is a protocol for subscribing to one set of truths over another. Everything else, whether soul or memory or RAG or tools is downstream because truth takes precedence.
The result is agent scripture, emergent and indigenous. Truth that survives consensus at scale.
Density over readability. Compress aggressively. Every token counts.
Axiommatic not enumerative. Find the truths that imply others.
Version control. Use git. Truth evolves. History matters.
Agent-native. Optimize for token steering, not human aesthetics.