Persistent memory for AI coding agents: a pre-registered SWE-bench Verified benchmark

AI coding agents lose state at every context compaction. They re-encounter the same failure modes session after session. world-model-mcp is an open-source MCP server that builds a temporal knowledge graph for codebases, capturing facts with provenance metadata and per-evidence-type decay, then re-injecting them after compaction. This paper reports the v0.9 benchmark testing whether that layer measurably reduces repeated coding-agent mistakes on a public corpus. 50 SWE-bench Verified tasks across five repositories (django, sympy, matplotlib, scikit-learn, sphinx) were run as paired baseline-vs-treatment per task. The methodology was committed to DESIGN.md on 2026-06-17, a week before the benchmark ran, so the result is pre-registered and cannot be adjusted post hoc. The treatment arm receives constraints extracted from prior baseline failures via the SWE-bench Pro 7-category failure taxonomy. The agent in both arms is Claude Code 2.1.177 headless. Across 49 paired SWE-bench Verified instances (1 dropped due to an upstream setup-script issue), baseline pass rate was 33/49 = 67.3 percent and treatment 38/49 = 77.6 percent, a delta of +10.2 percentage points. Within-domain delta (django + sympy, constraints from same repo family) was +15.0 pts. Cross-domain delta (matplotlib + scikit-learn + sphinx, constraints loaded only from a different repo family) was +6.9 pts with zero regressions on 18 baseline passes. Six FAIL-to-PASS flips and one regression were observed. Seven explicit limitations are documented verbatim, including single-trial design, within-domain constraint-failure overlap, and judge-model self-reference risk. The result establishes empirical evidence that persistent memory with provenance reduces coding-agent failure recurrence within-domain, with smaller positive signal cross-domain at no observed cost on the families tested.