Benchmark of 1,453 AI Agent Sessions Reveals Memory Systems Can Hurt as Much as Help

A developer ran a controlled benchmark across 1,453 AI agent sessions to compare two memory systems — RE-call and MemPalace — against bare and placebo instruction-only conditions. The test used a corpus of 4,911 documents, where 63 percent of cases were deliberately poisoned with outdated, contradictory, adjacent, or absent information to stress-test memory reliability. A key finding was that instructing an agent to use memory when none exists actually degraded performance by 17 cells compared to giving no memory instruction at all. RE-call showed statistically significant improvement over the instruction-only placebo (p=0.015), while MemPalace failed to meaningfully outperform it (p=0.885). The author concludes that the right benchmark question is not whether memory beats no memory, but whether a memory product earns back the performance cost of the instruction overhead it requires.