What is OIDA? Epistemic memory for organizational AI
OIDA is a way to organize an AI system's memory. It labels records as decisions, evidence, plans, or open questions, then connects them. These connections show which plan a decision replaced, where sources disagree, and what still needs an answer. Researchers call this epistemic memory.
When an old plan becomes the answer
Imagine asking an AI what your team decided. It finds an early proposal and a later decision, both about the same subject. To answer correctly, it needs to know which one still holds. Retrieval-augmented generation (RAG) gives an AI relevant passages to read, but ranking passages by similarity does not record that distinction.
Epistemic infrastructure stores the status alongside the record. A label says whether it is a decision, a plan, or an unanswered question. Connections show how it relates to other records. A priority score helps decide how much weight to give it when answering.
A typed, signed graph
A graph is a set of records and connections between them. The records are called nodes; the connections are edges. Each connection has a direction and a weight. SUPPORTS links evidence to a claim it strengthens; CONTRADICTS links claims that disagree. OIDA uses these connections, the records' labels, and their dates when choosing information for an answer.
Try it
Switch retrieval strategies and watch what each surfaces for the same question.
Agent query
“What is our position on NovaTech?”
Each node is a typed memory object — a decision, a plan, evidence, a hypothesis, or an open question — sized by its priority weight. Select a node to inspect it.
Choose a retrieval strategy. Semantic ranks by topical relevance; Epistemic (OIDA) also conditions on graph state — current vs. superseded, contradicted, and unresolved.
Epistemic graph for the query: What is our position on NovaTech?
Typed memory nodes:
- DECISION (priority 0.96): Co-invest $2M in NovaTech's seed round.
- PLAN (priority 0.22, superseded): Initial plan: lead a $5M seed round in NovaTech.
- CONSTRAINT (priority 0.90): Fund cannot hold more than 15% in a single seed deal.
- EVIDENCE (priority 0.78): Founding team: ex-Stripe engineers with strong technical pedigree.
- NARRATIVE (priority 0.70): Fund thesis: back technical founders building developer tooling.
- EVIDENCE (priority 0.74): TAM = $12B, per NovaTech's pitch deck.
- EVIDENCE (priority 0.76): TAM ≈ $3B, per the independent due-diligence dossier.
- HYPOTHESIS (priority 0.34): NovaTech can reach $10M ARR within 24 months.
- OBSERVATION (priority 0.38): A direct competitor raised a large round last month.
- QUESTION (priority 0.62): Unresolved: is NovaTech's core patent enforceable in the EU?
- EVALUATION (priority 0.54): Partner conviction score: 7/10.
Signed relations:
- Co-invest $2M SUPERSEDES Lead $5M seed (A replaces B; B is demoted, not deleted)
- ≤15% per deal BLOCKS Lead $5M seed (A actively prevents B)
- Founder pedigree SUPPORTS Co-invest $2M (A provides evidence strengthening B)
- Dev-tooling thesis SUPPORTS Co-invest $2M (A provides evidence strengthening B)
- TAM ~$3B (dossier) CONTRADICTS TAM $12B (deck) (A contradicts B)
- TAM $12B (deck) SUPPORTS $10M ARR / 24mo (A provides evidence strengthening B)
- Conviction 7/10 BASED_ON Founder pedigree (A is the logical grounding of B)
- EU patent enforceable? BLOCKS Co-invest $2M (A actively prevents B)
- Competitor raised SUPPORTS $10M ARR / 24mo (A provides evidence strengthening B)
Semantic retrieval: Ranks the passages most topically related to NovaTech — but the top hit is a superseded plan, and only the optimistic side of the market-size contradiction is surfaced. It misses: The current decision (D1) that superseded the $5M plan; The conflicting due-diligence TAM (E3) — only the deck's $12B appears; The open patent question (Q1).
Epistemic retrieval: Returns the current decision (not the superseded plan), flags both sides of the TAM contradiction, and surfaces the unresolved patent question as an open knowledge gap. It captures: The current decision D1 — the superseded plan P1 is demoted, not retrieved; Both sides of the TAM contradiction (E2 ↔ E3); The open knowledge gap Q1 (patent enforceability).
The nine node types
The nine labels help the system choose records for an answer. Each has a starting priority and a rule for how that priority changes over time. These are design choices, not a theory of how people think.
- DECISION1.00
Binding choice or commitment
Stable until superseded
- CONSTRAINT0.90
Hard structural boundary
Stable until changed
- EVIDENCE0.80
Verifiable supporting or refuting material
Slow decay
- NARRATIVE0.70
Persistent interpretive context
Stable contextual anchor
- PLAN0.65
Structured intention with horizon
Time-bounded decay
- EVALUATION0.55
Informed qualitative assessment
Moderate decay
- OBSERVATION0.40
Weak uninterpreted signal
Unreinforced decay
- HYPOTHESIS0.30
Unverified testable claim
Decay if untested
- QUESTION0.30
Open information need
Urgency grows while unresolved
Signed relations
When two records disagree, OIDA keeps both. It can reduce one record's priority or retrieve the pair together. A disagreement may need investigation; deleting a record would hide it.
| Relation | Coeff. | Semantics (A → B) |
|---|---|---|
| SUPPORTS | +1.0 | A provides evidence strengthening B |
| BASED_ON | +0.8 | A is the logical grounding of B |
| IMPLEMENTS | +0.7 | A operationally realizes B |
| SUPERSEDES | +0.6 | A replaces B; B is demoted, not deleted |
| REFINES | +0.5 | A narrows B without contradiction |
| DERIVES_FROM | +0.5 | A follows logically from B |
| ENABLES | +0.4 | A is a necessary condition for B |
| PRECEDES | +0.3 | A temporally precedes B |
| BLOCKS | -0.4 | A actively prevents B |
| CONTRADICTS | -0.6 | A contradicts B |
When is it worth using?
Before building a graph, check what your system needs to answer:
- Can the model read all the relevant documents at an acceptable cost?
- Does it need to distinguish decisions from plans, assumptions, and unanswered questions?
- How often do documents disagree or replace an earlier decision?
- Does an answer need to flag what is still unknown?
- Must the system follow a decision across several documents?
If the model can read the relevant documents affordably, start there. Ordinary RAG may be enough for a question answered by one source. A graph is worth considering when an answer depends on changes, disagreements, or unanswered questions across sources.
FAQ
- What is epistemic memory?
- Epistemic memory records the status of information alongside its contents. It can say that a record is a decision, evidence, a plan, a hypothesis, an observation, or an open question. It also records disagreements and which records were replaced. The AI can consult that status when answering instead of reconstructing it from the text each time.
- What problem does OIDA solve?
- An AI may find an old plan and report it as the current decision, quote only one side of a disagreement, or answer a question that is still unresolved. OIDA records the status and relationships of those sources so the system can use them when choosing information for an answer.
- Is OIDA a knowledge graph?
- Yes. A graph is a set of records, called nodes, and connections between them, called edges. OIDA labels each record by its role, such as decision or hypothesis, and stores dates, priorities, and relationships. A connection can show that two claims disagree or that one record replaces another. These details affect which records it retrieves.
- Do I need OIDA, or is RAG enough?
- If the model can read all the relevant documents at an acceptable cost, start with that. Ordinary RAG may be enough for a fact found in one source. Consider a graph when answers depend on changing decisions, conflicting sources, unanswered questions, sources with different authority, or connections across several documents within an input budget.