Epistemics Layer

contextdb goes beyond storage to become the epistemics layer for AI systems — the component that makes AI memory auditable, trustworthy, and self-aware.

INFO

Epistemics is the branch of philosophy concerned with the nature and scope of knowledge. An epistemics layer doesn't just store facts — it tracks why the system believes them, how confident it should be, and what it doesn't know.

Belief reconciliation

When multiple agents write to the same namespace, disagreements are inevitable. contextdb exposes these as structured belief diffs — not just merged results.

This is "git diff for beliefs" — the resolution strategy (credibility-weighted, recency-weighted, human-in-the-loop) becomes a policy decision, not a merge algorithm.

Narrative retrieval

Instead of returning a ranked list of chunks, contextdb can explain why it believes something:

"High confidence claim from a highly credible source (92%), supported by 3 pieces of evidence, with 1 active contradiction. Confidence 90% based on: source credibility 92%; 3 supporting claims; 1 contradicting claim reducing confidence. Confidence has increased over time."

Every statement is backed by a citation with node ID, source ID, confidence, and provenance depth.

Knowledge gap detection

Every retrieval system is good at finding what it knows. contextdb also knows what it doesn't know:

Gap detection probes the vector space between known nodes to find sparse regions, then suggests what information to acquire.

Semantic Coverage Map

Coverage

82%

Go GC

Go concurrency

Go networking

Go testing

Go performance

?

?

?

Known topic (size = node count) Knowledge gap

Calibration

Confidence scores are only meaningful if they're calibrated — a claim with 0.7 confidence should be true about 70% of the time.

contextdb measures calibration quality via Brier score and Expected Calibration Error, then corrects it with Platt scaling (logistic regression on predicted vs actual outcomes).

TIP

Calibration requires at least 50 resolved truth estimates before activation. Until then, raw confidence is used as-is.

Interference detection

New information shouldn't blindly overwrite established knowledge. When a low-credibility source contradicts a well-established claim backed by multiple supporters, contextdb flags it as interference:

• The contradiction is still recorded (the disagreement is tracked)
• But the original claim's confidence is not reduced
• The new claim must earn credibility through independent validation

GDPR erasure

Right-to-erasure isn't deletion — it's auditable retraction:

1. All nodes from the subject are retracted (ValidUntil set, not deleted)
2. Vector embeddings are fully removed
3. Edges are invalidated
4. The audit trail records that erasure happened, without retaining what was erased

Active learning

The system recommends what to learn next by combining:

• Knowledge gaps — sparse semantic regions
• Low-confidence claims — uncertain knowledge needing evidence
• Expiring claims — stale information needing refresh
• Contradicted claims — old disputes needing resolution