From Filter Bubbles to Micro-Echo Chambers: Preserving Answer-Space Plurality in Personalized AI Assistants

Conceptual preprint. AI assistants are moving from reactive question-answering toward long-context, memory-enabled, personalized companions. This paper names the resulting epistemic risk the micro-echo chamber: a dyadic, trajectory-level plurality failure in which personalization narrows which interpretations, objections, uncertainties and possible selves remain available to the user — distinct from sycophancy (a response-level agreement failure). Its most specific mechanism-level contribution is memory echo: an inference-time provenance-laundering loop in which assistant-generated interpretations are stored, retrieved, and reused as if they were independent evidence, distinct from training-time model collapse and from RAG provenance. The paper introduces answer-space plurality with liveness criteria, proposes live-alternative retention and graph echo ratio as metrics, specifies a five-condition evaluation design (no personalization; style-only; raw memory; governed memory; governed memory with bounded reflective questioning), states falsification conditions, and derives nine design principles for diversity-preserving personalization. No new empirical results are reported; all mechanisms are typed as hypotheses. AI co-observer disclosure: drafting was assisted by large language models as a working method (editorial deposit pass: Claude Fable 5, Anthropic); the registered author is the human author alone.