arXiv 'Engram': A Lean Retrieved Context Beats Full-History for Agent Memory — and Ships Open-Source

arXiv·medium signal

Engram, a bi-temporal memory engine, scored 83.6% vs 73.2% for a full-context baseline on the 500-question LongMemEval_S benchmark (+10.4 pts, statistically significant) while using ~9.6k tokens instead of 79k — roughly 8x fewer. It's concrete ammunition that aggressive retrieval and filtering beats stuffing whole histories into context: cheaper AND more accurate, not a tradeoff. Code, reproducible harness, and per-question logs are released CC-BY-4.0 at github.com/ly-wang19/engram (submitted ~June 5, not previously surfaced).