Deterministic Projection Memory for Enterprise AI Agents
A recent paper on arXiv (2604.20158) introduces Deterministic Projection Memory (DPM) aimed at enterprise AI agents operating in regulated sectors such as underwriting and tax audits. The researchers contend that traditional stateful memory frameworks compromise essential system attributes required for compliant use: deterministic replay, auditable reasoning, multi-tenant isolation, and statelessness for scalability. DPM employs an append-only event log combined with task-conditioned projections during decision-making. In experiments involving ten regulated decision scenarios across three memory budgets, DPM demonstrated comparable performance to summarization-based memory at higher budgets and surpassed it under significant compression: at a 20x ratio, factual accuracy increased by +0.52 (Cohen's h=1.17, p=0.0014) and reasoning coherence improved by +0.
Key facts
- Paper arXiv:2604.20158 proposes Deterministic Projection Memory (DPM) for enterprise AI agents.
- DPM addresses regulated domains such as underwriting, claims adjudication, and tax examination.
- Stateful memory architectures violate four systems properties: deterministic replay, auditable rationale, multi-tenant isolation, and statelessness.
- DPM consists of an append-only event log plus one task-conditioned projection at decision time.
- Tested on ten regulated decisioning cases at three memory budgets.
- At a 20x compression ratio, DPM improves factual precision by +0.52 (Cohen's h=1.17, p=0.0014).
- Reasoning coherence also improved at high compression.
- The paper argues regulated deployment is load-bearing on the four properties.
Entities
Institutions
- arXiv