Multi-Agent Memory Audit Methodology

A methodology for auditing and optimising memory configuration across a multi-agent AI system. Validated on a 13-agent production system running at 65% memory tax, meaning 65% of each agent context window was consumed by memory loading before any conversation started. Phase 1 alone reduced that to around 35%.

Memory tax is the core metric: what percentage of available context does memory loading consume at session start? Most operators do not know their number. Industry best practice is 20-30%.

The methodology runs in four phases. Phase 0 maps every agent file dependency before touching anything. Phase 1 is boilerplate extraction: identical configuration blocks duplicated across agents are extracted once and referenced, producing around 57% size reduction at zero risk. Phases 2 and 3 handle selective loading and archival.

What is included: the full methodology document, the Phase 0 audit template, the dependency mapping format, boilerplate extraction pattern with examples, the OpenClaw memory_search path issue, and the automated consolidation pipeline pattern.

Best for operators running 3+ agents who suspect their memory configuration is inefficient but do not know where to start.