Memory as Infrastructure: Reliability Engineering for Persistent Agent Memory in Months-Long LLM-Assisted Development

An instrumented N=1 operational record of SIx Harness: 78,933 hook invocations, 85 recorded failures (none silent), and seven design principles for keeping agent memory coherent across months of context compaction.

Mike Helwig· 2026· DOI 10.48550/arXiv.2609.05510

The core problem

LLM coding agents are crossing a threshold from task-scale to project-scale work: single assisted efforts that run for months, survive repeated context compactions, and operate on codebases far larger than any context window. This shift turns memory from a convenience feature into critical infrastructure, because an agent that forgets a decision, re-litigates a dead-end, or silently loses a convention will degrade the very project it is meant to accelerate. The paper reports operational experience from one such effort: a research project under a single continuous Claude Code session line since January 2026, driving a 633,000-line codebase, with its memory subsystem continuously instrumented since July 2026. From the need to keep this effort coherent emerged **SIx Harness**, an open-source memory and continuity infrastructure. The central claim is that the missing piece in operational practice is not better retrieval alone, but **reliability engineering for the memory subsystem itself**: a session-start health gate with discriminated failure modes, heartbeat telemetry designed so that no enumerated failure mode can pass unrecorded, and alert-fatigue budgeting borrowed from SRE prac

Innovation

The operational record is the paper's primary result. Across the instrumented period, the subsystem logged **78,933 hook invocations** and **85 recorded failures, none silent**. The temporal distribution is striking: **84 failures occurred in the subsystem's first three weeks**, **one since**, and **none in the final 20 days**. This pattern is consistent with a reliability-engineering curve in which early failures are surfaced, diagnosed, and structurally fixed rather than merely suppressed. The injection layer's **ten-day precision instrument shows zero false fires against an intact denominator**, indicating that precision-gated context injection did not degrade into over-injection during the measured window. The authors also report **three production incidents traced from instrument reading to structural fix**, demonstrating that the telemetry was actionable rather than merely descriptive. The heartbeat design is central to the no-silent-failure claim: because the absence of a heartbeat is itself an enumerated failure mode, a stalled or crashed memory subsystem cannot pass unrecorded. Alert-fatigue budgeting is reported as a borrowed SRE practice applied to keep the operator re

LLM coding agents are crossing a threshold from task-scale to project-scale work: single assisted efforts that run for months, survive repeated context compactions, and operate on codebases far larger than any context window. This shift turns memory from a convenience feature into critical infrastructure, because an agent that forgets a decision, re-litigates a dead-end, or silently loses a convention will degrade the very project it is meant to accelerate.

The paper reports operational experience from one such effort: a research project under a single continuous Claude Code session line since January 2026, driving a 633,000-line codebase, with its memory subsystem continuously instrumented since July 2026. From the need to keep this effort coherent emerged **SIx Harness**, an open-source memory and continuity infrastructure. The central claim is that the missing piece in operational practice is not better retrieval alone, but **reliability engineering for the memory subsystem itself**: a session-start health gate with discriminated failure modes, heartbeat telemetry designed so that no enumerated failure mode can pass unrecorded, and alert-fatigue budgeting borrowed from SRE practice.

Why it matters

The paper's central argument is that persistent agent memory should be treated as infrastructure, not as a feature. Once an LLM-assisted effort runs for months across repeated context compactions on a codebase far larger than any context window, the memory subsystem becomes a dependency whose failure modes are project-threatening. The authors argue that the operational practice gap is reliability engineering for that subsystem: discriminated failure modes, heartbeat telemetry that cannot fail silently, and alert-fatigue budgeting. From the record, the authors distill **seven design principles**. While the abstract does not enumerate them individually, the surrounding mechanisms imply their shape: local-first hybrid retrieval, precision-gated injection, anti-recurrence stores for decisions and dead-ends, compaction-surviving conventions, a session-start health gate with discriminated failure modes, heartbeat telemetry designed against silent failure, and alert-fatigue budgeting. The temporal pattern of failures—84 in the first three weeks, one since, none in the final 20 days—supports the view that early instrumentation plus structural fixes produces a stabilizing reliability curve. The limitations are stated plainly: N=1, no control arm, self-reported. These constraints mean the record is existence proof and hypothesis generation rather than confirmatory evidence. To address this, the authors publish a tagged pre-registered ablation protocol that any team can run with the released MIT-licensed kit, inviting the community to test the design principles under controlled conditions. The broader implication is that as LLM coding agents move from task-scale to project-scale, the differentiator may be less about model capability and more about the reliability engineering wrapped around persistent memory. The paper's contribution is to make that infrastructure visible, instrumented, and reproducible.

Who should read this

CS practitioners and researchers

Opening member content…