What question did this study set out to answer?

To address the staleness problem in persistent memory used by coding agents in multi-agent environments.

July 9, 2026Open Access

Memory Is Not Live State — Cross-Agent Memory Governance by Transplanting the Reflective Context Governor onto Agent Memory (memgov v0.1.0)

Key Points

To address the staleness problem in persistent memory used by coding agents in multi-agent environments.
Documented a near-miss incident involving memory staleness.
Analyzed root causes including store fragmentation and guard misconfigurations.
Deployed a cross-agent shared state system with freshness monitoring and conflict resolution mechanisms.
Identified 17 suspected memory drifts on the first deployment day.
Detected true positives including three misremembered unpublished papers.
Validated memory governance mechanism through deliberate screening and monitoring.

Abstract

Coding agents ship with persistent memory, and persistent memory has a structural defect: a memory is a snapshot of the world at write time, while the world keeps moving. Status-bearing claims ("not yet deposited", "awaiting review", "tool X is broken") are falsified the moment another session, another agent, or the human owner changes the world — and in multi-agent workspaces this staleness is the steady state, not the edge case. This systems paper documents a real near-miss (an agent's memory asserting a published preprint was still undeposited, one day after publication), the root-cause analysis (store fragmentation, a guard watching the wrong index, cross-agent staleness), and the countermeasure: a same-day transplant of the Mobius Reflective Context Governor's robust core — readiness before injection, committed versus proposed authority, non-compensatory gates, governed friction, visible exclusion — from the context-window domain onto the memory domain, under the axiom RelyOnMemoryₜ ⇒ MemoryReadyₜ. The deployed system (a cross-agent SHAREDSTATE commitment surface, a deterministic drift scan producing CONFLICTMAP review queues, freshness heartbeats, and an append-only measurement ledger) recorded a full detect-verify-repair cycle on its first day: 17 suspected drifts, true positives including three papers misremembered as unpublished and a toolchain misremembered as broken. The mechanism is generalized and released as memgov v0. 1. 0 (AGPL-3. 0-or-later, stdlib-only Python, adapters for Claude Code and Codex, whose session-start hooks share one JSON contract). The paper includes a full exposition of the RCGov theory being transplanted, an explicit concept-to-mechanism mapping, an evaluation plan with pre-stated falsification criteria, and a scope firewall: no benchmark superiority is claimed, detection is deliberate regex-level screening, and engineering success validates nothing upstream of the engineering. AI co-observer: Claude Fable 5 (Anthropic) served as drafting partner and as the operating agent in the documented incident, deployment, and release; the registered author and responsible party is the human author alone.

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper