Facts: Curated Knowledge for Humans and Agents

Facts is a knowledge management system focused on determining which information should become accepted knowledge, when, and how. It emphasizes multi-player consensus alongside rapid capture and recall capabilities. The system's primary unit is a proposition, a statement that can be assessed as true or false. While the proposition itself does not automatically become a fact, it can participate in the ledger's accepted knowledge once validated by someone or something.

This subtle distinction is crucial, as it enables multi-player curation and offers transparency, traceability, and accountability. Facts draws inspiration from Git-like properties, ensuring that the old fact persists even after edits, maintaining a history of how and why knowledge changes over time. This allows for a clear distinction between considering a proposition and believing it to be true.

Facts operates agnostic of whether the actor is a human, an agent, a service, or an organization. It establishes an operating environment for a registered actor, complete with a public/private key pair and permissions to engage in the system. The human can assign a name to the actor, which is important for identifying the entity responsible for the actions taken.

The common agent-memory loop involves an agent proposing new information and revisions, while a mistake could potentially be committed to memory as truth, leading to potential catastrophic failures if shared among agents. Facts addresses this by introducing a decision boundary, allowing an agent to learn and propose new information without immediately seeing it as accepted knowledge.

Tagging facilitates organization and narrows vector search results, enabling more structured knowledge retrieval without becoming a heavy-duty knowledge system. Agents can access accepted facts to retrieve relevant information, such as policies and their derivations, without needing the entire architectural context.

The system's implementation allows various applications to benefit from its knowledge management capabilities. For instance, an engineering agent working on billing can retrieve both the policy and the history of how it was established. Incident response becomes more efficient, with AI agents accessing the same incident response ledger to analyze customer tickets and determine actions for accepted or rejected propositions.

Facts' principles are versatile and can support diverse systems, fostering collaboration across teams working from a shared knowledge domain. Agents can be granted access only to the knowledge domains relevant to their tasks, ensuring that the ledger remains a durable project memory even after the project concludes. The knowledge itself acts as both communication and a record, with each agent responsible for its own work and memory, without Facts dictating an agent framework.

The system's design emphasizes the importance of consensus in determining whether a proposition becomes effective. Different models or agents can be combined, reducing correlated failure modes. Agents have permission to propose new information but not to decide on its acceptance. Humans step in only when there is contention, managing exceptions rather than reviewing every memory update.

The key takeaway is that Facts provides a structured approach to knowledge management, enabling efficient retrieval of relevant information while maintaining a clear history of knowledge evolution.