Algorithmic Psychodynamics in Multi-Agent Memory: Modeling Defense Mechanisms, Epistemic Dissonance, and Supersession Governance

Persistent memory in multi-agent AI systems is increasingly a problem of epistemic governance, not solely retrieval. When autonomous agents exchange assertions, they may propagate stale, contradictory, or adversarially contaminated information into subsequent decisions. This paper proposes Algorithmic Psychodynamics as a design vocabulary that draws functional analogies from cognitive dissonance, defensive regulation, and memory reconsolidation without attributing consciousness or clinical states to software. We define an epistemic belief tuple, separate contradiction detection from confidence-weighted arbitration, introduce a guarded handoff protocol, and describe semantic quarantine, action-authority isolation, temporal supersession, and provenance-aware retraction. A compact JSON-LD example illustrates interoperable representation using W3C PROV-O relations. Comparison with recent work on stale-agent memories and governed shared memory narrows the proposed contribution to an integrated conceptual design rather than a demonstrated new algorithm. We specify failure conditions and a falsifiable evaluation agenda; no prototype performance or experimental effectiveness is claimed.