Retrieval Is Not Enough: Why Organizational AI Needs Epistemic Infrastructure

Imagine asking an AI to summarize what your team decided. It finds an early proposal, a later decision, and a market update that challenges the plan. All three concern the same subject. The answer depends on knowing which record still holds and how the others relate to it.

When relevant is wrong

In this example, the decision replaced the proposal. The market update raises a disagreement, and a regulatory question remains unanswered. If the AI treats all four records as equivalent evidence, it can give a well-written answer that cites real documents but misreports the team's current position.

The model may have quoted its sources accurately. The mistake is treating an old plan as a current decision.

The failure is representational

The records need labels and connections that show their status. An organization's documents contain more than facts. They contain decisions, constraints, plans, hypotheses, evidence, observations, contradictions, and unresolved questions. An AI answering a question needs those distinctions to decide which records to use.

Vector retrieval searches for passages similar in meaning to a question. A reranker reviews the candidates to choose the most relevant ones. Neither step, by itself, stores which decision replaced which plan. If that information appears only in prose, the model must work it out again each time from the documents it receives.

Epistemic infrastructure

A memory layer can store the records' status alongside their content. This is what we mean by epistemic state: what is decided, disputed, replaced, or still unknown. In the OIDA paper, we formalize this as a typed, attributed, signed, time-indexed graph. Each node is a typed memory object with an epistemic role and a priority weight; edges encode support, dependency, contradiction, and supersession. Retrieval then conditions semantic similarity on graph state — so a superseded plan is demoted rather than surfaced, both sides of a contradiction can be retrieved together, and unanswered questions get their own records, which the system can retrieve.

If you want the concepts without the math, start with the OIDA explainer.

Reading the evidence honestly

In the pilot, when the entire corpus fits in the model's context window, the full-context baseline wins aggregate answer quality — it simply sees everything. In that setting, structured retrieval uses roughly 42 times fewer input tokens, and it surfaces explicit knowledge gaps far more consistently (ten times out of ten, versus five).

The paper argues for storing the status of records in memory. The pilot does not establish that a graph beats full context in general, or that the results hold in production. Further tests must isolate the contribution of changing priorities, check the accuracy of extracted connections, and measure behavior over time.

You can inspect the methods and released document collections in the paper.