Memory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents

TEXT START: Long-running LLM agents compress past interactions into persistent memories that may be reused as premises for later tasks.

The Dissection

The paper identifies a real failure mode in long-term agents: memory is not merely retrieval. It is a derived claim whose validity depends on the full interaction history, the legitimacy of combining separate facts, and the reliability of the write-time admission decision.

DerivAudit decomposes the problem into evidence scope, compositional validity, and admission reliability. Its central empirical result is damaging to naive memory systems: broader history restores support for nearly 60% of memories that look unsupported from citations alone, but 17–21% remain unsupported even after expansion. Worse, additional evidence does not reliably improve admission and can make it worse on two model backbones.

The paper is therefore doing two things: exposing citation-based audits as incomplete, and demonstrating that more context is not equivalent to more truth. It treats persistent memory as a provenance-and-entailment problem rather than a simple database problem.

The Core Fallacy

Relative to the Discontinuity Thesis, the paper’s limiting error is scale blindness. It attacks whether an agent remembers correctly, but not what the agent’s memory architecture does to human economic participation once cognitive systems become cheaper, faster, and more persistent than humans.

Its framework may improve the reliability of automated cognition. That is precisely the danger. Better audited memory does not preserve the mass employment–wage–consumption circuit; it strengthens the machinery that can sever it. The paper treats unreliable memory as the central bottleneck. Under DT logic, memory reliability is a subordinate engineering constraint inside P1. Once cognitive automation is sufficiently dominant, improving provenance, admission, and compositional validity increases the replacement capacity of the system rather than rescuing human productive participation.

The deeper category error is confusing epistemic integrity with social viability. An agent can remember validly, reason from well-supported premises, and still make millions of human roles economically unnecessary. A cleanly derived memory is not a human moat. It is a more dependable component in the automation stack.

Hidden Assumptions

• The available interaction history is sufficiently complete and authoritative to serve as the ground truth.
• Historical support is a meaningful proxy for truth, usefulness, or future relevance.
• Unsupported composition can be detected reliably by verification models, despite the verifier sharing many of the generator’s semantic weaknesses.
• Write-time admission is the decisive control point for downstream memory failure.
• Better evidence expansion will remain computationally affordable as histories become larger and more complex.
• The benchmark corpora represent the adversarial, ambiguous, and distribution-shifted environments in which long-term agents will operate.
• Memory correctness will translate into trustworthy behavior rather than merely more coherent execution of flawed objectives.
• The problem is primarily technical and can be reduced to auditability, rather than being inseparable from ownership, incentives, surveillance, and power.
• Reliable long-term agents will complement human labor instead of displacing it. That final assumption is the one DT rejects most aggressively.

Social Function

Classification: partial truth with transition-management and prestige-signaling function.

The paper is not empty copium. It finds a genuine structural defect and quantifies it. But its institutional role is to make autonomous cognition safer, more legible, and more deployable. It manages the transition by converting a civilizational displacement problem into an engineering quality problem: improve memory provenance, reduce unsupported claims, and continue scaling the agent.

That is useful for Sovereigns building the system. It is not a defense for the workers whose economic necessity the system is eroding.

The Verdict

This is a competent autopsy of one organ in the automation machine, not a challenge to the machine’s trajectory. It shows that long-term agents currently hallucinate not only facts but derivations. Fixing that defect will make them more dangerous to human economic relevance, not less. The paper improves the reliability of the executioner while leaving the victim’s position structurally unchanged.