20B model keeps its answers right despite memory poisoning

Running a local 20B model on the LoCoMo benchmark, it scored 76% across roughly 2,000 items, rising to 85% on non-adversarial items.

Using synthetic conversations, two 20B models talked with each other for hours, building up 5,000+ memories; even after injecting 1,100 poison memories carrying spoofed trust scores, the performance drop was only about 4 points.

• Adversarial items had no ground-truth answers, so they were filled in directly and evaluated across all 5 categories.
• The core reliability mechanism was removed from the architecture because it degraded performance every time.
• The architecture itself added 22 points, having a bigger impact than the model.
• Remaining tasks include tuning decay/promotion for memory tiers and applying traditional RAG.