Agent Memory Curator Agent — Development skill for Claude Code

Self-Evolving Agents Hallucinate at Scale: A Provable Memory Curator Architecture

**Mason Lee** Appbridge Inc. (Agentlas) [agentlas.cloud](https://agentlas.cloud) · appbridge@appbridge.co.kr

**Version:** v2.0 preprint draft (May 26, 2026) **Code:** [github.com/agentlas-ai/agent_memory_curator_agent](https://github.com/agentlas-ai/agent_memory_curator_agent) **License:** MIT (code), CC-BY 4.0 (paper)

---

Abstract

**We prove that self-evolving LLM agents — Hermes, Voyager-style skill libraries, and their production kin — accumulate retrieval hallucinations at a rate that approaches certainty as deployment horizons grow, unless durable memory writes are explicitly governed.** Monte Carlo simulation over 200 seeds confirms the bound: an uncurated self-evolving agent operating at 10 events/day with realistic 10% per-event hallucination probability reaches a **98.6% per-retrieval hallucination probability** within one year. A curator with admission rate $\alpha = 0.3$ and filter accuracy $\eta = 0.9$ cuts this to **30.7%** — a 3.2× reduction. A stricter curator ($\alpha = 0.5$, $\eta = 0.95$) cuts it to **17.7%** — a 5.6× reduction.

We formalize the *memory admission problem* in multi-agent settings and prove a *compounding-bound theorem*: under mild assumptions, the per-retrieval hallucination probability in an uncurated regime grows asymptotically as $1 - (\lambda t)^{-r_0 h_e}$, whereas a curator bounds the rate at $1 - (\alpha \lambda t)^{-r_0 h_e (1-\eta)}$, with the ratio diverging to infinity as $t \to \infty$ for any $\eta > 0$.

The bound motivates a system design. We propose the **Memory Curator Agent**: a dedicated specialist that owns durable memory writes while worker agents emit *structured memory events*. Memory is partitioned into four scopes — *agent repo*, *agent team*, *project*, *session* — corresponding to specialist, transacti