Cryptographic Provenance Gates Mitigate Multi-Session Memory Poisoning in Autonomous Agents

Autonomous agent architectures deploying persistent Retrieval-Augmented Generation (RAG) commit interaction history into long-term vector and relational storage. In extended multi-session deployments, unconstrained memory write policies introduce recursive self-poisoning: transient model confabulations emitted in Session t are retrieved in Session t+k as authoritative historical ground truth, compounding hallucinations into permanent systemic errors. In this work, we evaluate the mitigation of self-poisoning via an immutable Cryptographic Provenance Gate enforced through SQLite database triggers (Mem=1, Gnd=0 for assistant utterances). Across a pre-registered benchmark of N=100 paired multi-session probes (prereg-poisoning-v1 / v2) replicated across four Gemma-4 parameter scales (26B-A4B, 12B, E4B, E2B), the Provenance Gate achieves a 0.0% self-poisoning compounding rate compared to 100.0% for Naive RAG (p = 6.98e-10) while preserving 96.7%-100.0% conversational continuity, and reduces uncorroborated user falsehood compounding to 0/10 on 26B-A4B, 12B, and E4B under live corroboration verification (WO-39).