What Is Epistemic Memory? A Typed-Graph Alternative to Vector Retrieval

Epistemic memory records the status of information an AI uses. A record can be a decision, an untested hypothesis, or an unanswered question. Connections show whether another record replaced it or disagrees with it. The system stores this information so it can use it when answering, without reconstructing every relationship from the document text.

Why vector retrieval is not enough

Vector retrieval finds text similar in meaning to a question. Consider an example: two passages discuss an acquisition, but one is a proposal and the other records the decision that rejected it. Both are relevant. A similarity score alone does not record their relationship. A model that reads the documents may infer it, but must do so again for each answer. In the reported pilot, reading the full collection gave better overall answer quality.

Epistemic memory makes the status explicit and durable.

A typed, signed graph

OIDA represents this information as a graph. Nodes are records with labels. Edges are connections between records, with a direction and a weight. the OIDA architecture uses nine node types — decision, constraint, evidence, narrative, plan, evaluation, observation, hypothesis, and open question — each with an initial priority weight and a temporal behavior. A decision stays binding until something supersedes it; an open question grows more urgent the longer it stays unanswered.

Edges carry sign. Positive edges like SUPPORTS and IMPLEMENTS reinforce. Negative edges like CONTRADICTS and BLOCKS mark tension that topical similarity would otherwise hide. When two records disagree, OIDA keeps both. It can reduce one record's priority or retrieve the pair. A disagreement may need investigation, and deleting a source would hide it.

How retrieval changes

In an epistemic memory, ranking combines semantic similarity with graph state. A node's score depends on its content's relevance and on its type, its priority weight, and its local edges. The practical effect: ask "what did we decide?" and you get the current decision, not the proposal it replaced. Ask about a contested number and you get both sides of the contradiction, flagged. Ask an open-ended question and the system can tell you what is still unknown, rather than papering over the gap.

When it matters

Building and maintaining the graph has a cost. If your relevant corpus fits affordably in context, use the context. If your queries are mostly single-hop factual lookups, ordinary retrieval is enough. Consider it when decisions change, sources disagree, unanswered questions matter, or an answer must connect several documents within an input budget.

For the full treatment — the formal model, the released evaluation corpora, and the pilot's results and limitations — read Retrieval Is Not Enough.