When Memory Leaks: RAG-Induced Ambiguity and Fracture-Repair Hallucination

Frontier LLMs are exhibiting a new manifestation of hallucination through memory leakage in deployed systems. We document this phenomenon through systematic observation across four models and propose interventions with proof-of-concept validation. Large-scale validation remains future work.

Prior work established that hallucinations follow a fracture-repair process: ambiguity produces semantic fractures that models repair using proximate context (Evans, 2024). This framework successfully explained entity-tracking failures but could not account for why hallucinations persist despite explicit source-grounding.

This paper identifies memory leakage as a new ambiguity generator. Across four frontier models (GPT-5.2, Claude Sonnet 4.5, Grok 4.1, Gemini 3.0 Thinking Mode), we observe:

1. Memory leakage: RAG systems retrieve contextually inappropriate user history due to lack of provenance governance

2. Ambiguity generation: Leaked memory conflicts with source material, creating semantic fractures

3. RLHF-forced integration: Alignment incentives require models to synthesize conflicting inputs rather than acknowledge insufficiency

This triggers the established fracture-repair mechanism, but with a critical difference: repair material comes from leaked cross-conversation memory rather than source-proximate context. The result is complete, personalized outputs that are maximally persuasive because they deploy the user's own conceptual frameworks—but in inappropriate contexts.

In memory-enabled systems, memory-leakage hallucinations are not mitigated by source-grounding or improved retrieval quality. Instead, these measures often intensify the failure by increasing confidence in continuation while leaving authority conflict unresolved.The failure is architectural: expanded memory capacity without semantic governance creates retrieval ambiguity, while RLHF training prevents epistemic restraint. Solutions require RAG governance (provenance hierarchy, authority assignment, boundary enforcement) and significance-based filtering.

This work completes the fracture-repair framework by documenting how ungoverned memory expansion in RAG systems creates systematic ambiguity that triggers hallucination through established mechanisms.

This paper identifies memory leakage in retrieval-augmented generation systems as a previously undocumented trigger for hallucination. When cross-conversation user memory is retrieved without provenance hierarchy or authority assignment, it conflicts with current source material and creates ambiguity of authority rather than ambiguity of meaning. Across four frontier models, this conflict is recognized prior to generation but cannot be acknowledged due to alignment incentives that penalize uncertainty. Generation therefore proceeds by repairing the fracture using the most salient available anchor—typically the user’s own prior frameworks—producing complete, confident, and highly persuasive hallucinations. Low-substance source material amplifies this failure by weakening source anchors but is not causative. The findings demonstrate that hallucination prevention requires governance, not smarter models: specifically, retrieval provenance controls, authority assignment, and permission for epistemic stopping.

Withdrawal of Prior Framing

Earlier drafts of this work attributed the observed hallucination behavior to semantic insufficiency in source material and referred to this effect as The Pearl. Further investigation demonstrates that this framing was incomplete. Semantic insufficiency is neither necessary nor sufficient to trigger the phenomenon. We therefore withdraw both the name and the explanatory claim.

The underlying trigger is memory leakage: the retrieval of contextually inappropriate user history without provenance or authority governance. Semantic insufficiency acts only as an amplifier once memory leakage has occurred, increasing salience imbalance between retrieved memory and source material. The phenomenon documented in this paper is therefore not an effect of empty content, but of authority conflict induced by ungoverned memory retrieval.