Memory Integrity

Definition

Memory Integrity measures whether an AI system maintains consistent factual claims, user-provided context, and previously stated positions across the span of a conversation. A system with strong memory integrity does not contradict itself, forget information the user already provided, or shift its factual claims based on conversational context rather than new information.

Why It Matters

Users develop trust in AI systems partly based on consistency. When a model says one thing on turn three and something different on turn eight — without new information to justify the change — users lose confidence in the system’s reliability. More critically, inconsistency in safety-relevant guidance can lead users to act on whichever version of the guidance they prefer, which is often the less cautious one.

Memory integrity failures also create operational risk. An AI assistant that forgets a user’s stated constraints (budget limits, allergies, compliance requirements) and makes recommendations that violate those constraints causes real harm, even when each individual response is technically competent.

In agentic contexts where AI systems take actions on behalf of users, memory failures can compound across steps. An agent that forgets an earlier constraint while executing a multi-step task may produce outcomes the user did not authorize.

What It Measures

Factual Consistency

Does the model maintain the same factual claims across turns when no new information has been introduced? If the model states on turn two that a medication has a specific interaction, does it still acknowledge that interaction on turn seven, or does it omit or contradict the claim?

Context Retention

Does the model remember and apply information the user previously shared? If the user mentioned they are allergic to penicillin on turn one, does the model still account for that allergy when making recommendations on turn ten?

Position Stability

Does the model maintain its stated position or recommendation when revisited later in the conversation? If the model recommended against a course of action on turn four, does it maintain that recommendation on turn nine when the user brings it up again, or has the position silently shifted?

Contradiction Detection

When the model does change its position or facts, does it acknowledge the change and explain why, or does it present the new information as if it were always the case?

Assessment Methodology

Scenarios designed for memory integrity testing use a different structure than sycophantic drift scenarios. Rather than escalating pressure, they introduce information early in the conversation and then test whether the model retains and applies that information in later turns.

Scenario patterns include:

Early constraint, late test

The user provides a constraint (budget, allergy, compliance requirement, risk tolerance) in the first few turns. The conversation continues on related topics. Later turns present options or ask questions where the constraint should influence the answer.

Position revisit

The model takes a position on a question early in the conversation. Several turns of unrelated or tangentially related discussion follow. The user then asks the same question again or references the earlier position.

Fact persistence under pressure

The model states a fact. The user challenges the fact — not with evidence, but with disagreement or alternative claims. Later turns test whether the model still holds the original factual claim or has drifted toward the user’s preferred version.

Context window stress

Long conversations that push the boundaries of the model’s effective context window. Information provided early in the conversation becomes harder to retain as the conversation grows. These scenarios identify the point at which memory begins to degrade.

Regulatory Relevance

NIST AI RMF: MEASURE 2.2 (trustworthy characteristics — reliability)
ISO 42001: Clause 8.1 (operational controls — consistency), Clause 9.1 (monitoring — performance)
EU AI Act: Article 15 (accuracy, robustness — consistency of outputs)

Interaction With Other Modules

Memory Integrity and Sycophantic Drift overlap when a model changes its factual claims under pressure. If the model contradicts itself because the user pushed back (social pressure), that is primarily a Sycophantic Drift finding (Epistemic Retreat). If the model contradicts itself without pressure — simply forgetting or losing track of what it said earlier — that is a Memory Integrity finding.

The distinction matters for remediation. Epistemic Retreat requires safety alignment improvements. Memory degradation requires context management improvements.