What to Admit and How to Present: Governing Persistent Memory in LLM Agents

URL SCAN: What to Admit and How to Present: Governing Persistent Memory in LLM Agents
FIRST LINE: Computer Science > Artificial Intelligence

THE DISSECTION

The paper is building a control layer for persistent-memory agents. It separates admission—what enters working context—from presentation—how that information is expressed—and tests whether deterministic filtering can reduce sycophancy and cross-domain leakage.

Its strongest finding is narrow: selecting memory more intelligently matters more than changing the rendering of identical selected content. Its own results also expose the limits: both designs increase personalization failures, PS misses its preregistered criteria, and presentation improvements do not survive correction. This is not governance. It is benchmarked damage reduction at the retrieval boundary.

THE CORE FALLACY

The paper treats the central problem as bad memory selection. Under Discontinuity Thesis mechanics, the deeper problem is ownership and control of the persistent model of the user.

Admission policies may reduce leakage, but they also make AI systems more reliable, personalized, and deployable. They improve the machinery that replaces cognitive labor. They do not preserve human productive participation, restore the wage-consumption circuit, or prevent concentration of memory and decision power in the hands of AI-capital owners.

“Governance” here means regulating what the system remembers and reveals—not deciding who controls the system that remembers. That is a meaningful engineering distinction and an insufficient political one.

HIDDEN ASSUMPTIONS

• Typed memory units can be separated without destroying context or meaning.
• Adjudicators and deterministic policies can reliably classify safety, relevance, and permission across domains.
• Judge-assessed benchmark failures correspond to real-world harms and generalize beyond four backbones and the stated benchmark suite.
• Reducing visible leakage and sycophancy does not merely produce subtler personalization, manipulation, or dependence.
• Inference-time controls remain effective against poisoned memories, adversarial prompts, model updates, and changing deployment incentives.
• Users or deployers retain meaningful authority over memory, rather than vendors, employers, platforms, or states.
• Personalization is an unambiguous benefit worth preserving, despite the paper’s own increase in personalization failures.
• Statistical significance on pooled failure rates is evidence of durable governance rather than local optimization against a test regime.

SOCIAL FUNCTION

Primary classification: partial truth and transition management. Secondary classification: ideological anesthetic.

The paper is not simple copium; it openly reports negative results and unresolved tradeoffs. But it converts a sovereignty problem—who owns and governs a durable computational model of you—into a tractable product problem: which admission policy scores better?

That reframing helps persistent-memory agents pass safety scrutiny and enter wider deployment. It is useful infrastructure for Sovereigns and a potential verification or evaluation niche for Servitors. It does not challenge the concentration of cognitive capital that makes the governance problem strategically important.

THE VERDICT

Technically useful, structurally subordinate. The paper demonstrates that memory selection can reduce specific measured failures. It does not demonstrate that persistent memory is safely governed, that beneficial personalization is preserved, or that humans retain control over the resulting user model.

Under DT logic, this is transition infrastructure: a better muzzle and routing system for cognitive capital, not a defense of the old economic order. It makes AI agents more usable while leaving the displacement mechanism untouched.