The Memory-Citation Feedback Loop: How Conversational Context Creates Per-User Citation Advantages in AI Systems

1. Introduction

Modern AI conversational systems increasingly incorporate persistent memory features that retain information across sessions. When a user extensively discusses their business with an AI system, that business information becomes part of the conversational context available during subsequent interactions. A natural question arises: does this accumulated memory create a citation advantage for the user's business when the AI system generates answers to related queries?

If memory influences citation behaviour, this would represent a previously undocumented feedback loop: using an AI system extensively about one's own business could create a per-user advantage in AI-generated recommendations, independent of any search optimisation efforts. This paper presents preliminary observational evidence testing this hypothesis.

2. Methodology

2.1 Test Design

Two tests were conducted in a memory-enriched AI conversation where extensive business information had accumulated through prior discussion.

Control test: A query in a domain where the user's business had no relevance. This tested whether memory inappropriately injects entities where they are irrelevant.

Treatment test: A query in a domain where the user's business was relevant but did not appear in search results. This tested whether memory supplements search results with entities from memory during answer synthesis.

2.2 Measurement

For each test, we recorded: whether the user's entity appeared in search results, whether the user's entity appeared in the generated answer, what specific information about the entity was included, and the apparent source of that information (search results vs. memory).

3. Results

3.1 Control Test

In the control query (a domain where the user's business had no relevance), the user's entity did not appear in search results and was not mentioned in the generated answer. No memory leakage was detected. This confirms that memory does not indiscriminately inject known entities into unrelated queries.

3.2 Treatment Test

In the treatment query (a domain where the user's business was relevant), the user's entity did not appear in any search results. However, the generated answer included the entity with specific details drawn from memory: pricing tiers, service model description, team structure, and competitive positioning. Eight competitor entities were found through search; the user's entity was the only one added from memory.

3.3 Two-Layer Model

3.4 Magnitude of Advantage

The memory-derived citation advantage consisted of approximately one additional entity mention plus specific details (pricing, service model, competitive positioning) that search-only competitors did not receive. The advantage existed only in conversations with the memory-enriched user and would not be present for a different user submitting the identical query.

3.5 Mechanism

The AI system's memory is injected into the context window before the conversation begins. When search results are returned, the model evaluates them with memory context already loaded. This causes three observable effects: (1) recognition of relevance that might otherwise be missed, (2) addition of supplementary information that search did not surface, and (3) richer framing of known entities compared to search-only competitors.

4. Discussion

The per-user citation feedback loop raises several important considerations. First, it means that extensive AI system usage about one's own business constitutes an inadvertent form of GEO optimisation -- but only for one's own conversations. The advantage does not extend to how the AI system treats the business when other users ask about it.

Second, the two-layer model suggests that search retrieval fairness is maintained (all users see the same search results) while answer synthesis fairness is not (users with richer memory context receive richer answers about entities they have discussed). Whether this constitutes a problematic bias or a legitimate personalisation feature is an open question.

Third, the long-term effects on model training are unknown. If many users discuss a business extensively with the AI system over time, aggregate conversation data could theoretically influence future training data, creating an indirect feedback loop that extends beyond per-user effects. This remains speculative.

5. Limitations

• N=2: Only two tests (one control, one treatment). The observation requires replication with larger sample sizes.
• Single AI system: Only one AI platform's memory system was tested. Other platforms may implement memory differently.
• Introspective framework: The two-layer model is an interpretive framework, not a confirmed architectural description.
• No fresh-conversation replication: The treatment test was conducted in a memory-enriched conversation with prior discussion context, making it impossible to isolate memory effects from conversational priming effects.
• Confound: conversational priming: Extended conversation about the user's business may create a priming effect separate from persistent memory.

6. Conclusions

We observe preliminary evidence for a per-user citation feedback loop operating at the answer synthesis layer but not the search retrieval layer. An entity absent from search results appeared in the generated answer for a memory-enriched user, with specific details drawn from conversational memory. The effect is per-user only and does not affect other users' experiences. This observation (N=2) requires replication across larger samples, fresh conversations (isolating memory from priming), and multiple AI platforms.

Confidence: LOW. N=2 tests. The two-layer model is an introspective framework. Fresh-conversation replication not yet performed.