Memory Echo and How to Stop It: Provenance-Tagged Memory with Deterministic Output-Time Self-Citation Verification — Systems Paper — the MOBIUS-RQA Governance Primitives for Assistants with Long-Term Memory, and the Bounded Reflective Questioning System They Keep Honest

Assistants with long-term memory share a failure mode this project names memory echo: an interpretation the model produced on an earlier turn is stored, retrieved later, and treated as independent, externally grounded evidence. Its terminal forms are a self-reinforcing echo chamber (the assistant converging on its own guesses) and the fabricated self-citation — "as I established earlier, X" — where X was never established. We describe MOBIUS-RQA, an open-source (AGPL-3.0) system whose memory layer answers this with two deliberately boring, liftable primitives. (1) Provenance-tagged memory: every stored node carries its source class (user-said / assistant-generated / external / system), and the tag travels with the node into every assembled context, so an assistant-generated item can never silently present as independent evidence. (2) Deterministic output-time self-citation verification: before a response is emitted, every memory reference it makes is checked in code against the set of node identities actually injected this turn; references to anything else are fabricated self-citations and are stripped. The model is not consulted about its own provenance — the verifier is deterministic, not a self-report. A published pilot measures the failure and the remedy: ungoverned raw memory collapses answer-space plurality to roughly one seventh of the memoryless baseline, governed memory restores it, the output-time verifier removes fabricated self-citations deterministically, and the pattern reproduces across six local models in three families (Gemma-4, Llama-3.1, Qwen-3.5/3.6). We then describe the larger system these primitives keep honest: a bounded reflective questioning assistant — one that structures tensions, excavates assumptions, and produces deeper questions rather than answers — with a staged response surface (fixed reply → direct answer → short-then-full two-stage output → instrument panel) and explicit reflection budgets. The memory governance is what entitles the reflection: a system whose trade is questions must not build them on its own recycled guesses. Scope is stated plainly: the pilot establishes the mechanism at small n; production use wants larger n and independent integration. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.