A new revision of an arXiv paper turns AI memory into a trust problem, not just a storage one.
The paper, arXiv:2501.09166, originally posted in January 2025, proposed a Retention Layer that lets a Transformer read and write persistent memory as it works. This revision reframes the hard part: deciding what to keep is treated as a social learning problem, covering when to trust observed behavior, whom to learn from, and how much independent agreement should be required before something sticks. The authors add gating for encoding based on surprise, outcomes, and earned credibility, a credibility-weighted quorum for consolidating memories, and reconsolidation driven by how well a memory holds up when reproduced. Testing with a long-running assistant led to two added rules: a model's own outputs cannot count as supporting evidence, and a user's statement is kept after a single mention.
In simulations involving world drift and three memory-poisoning attacks, the new lifecycle held accuracy between 98.9 and 99.6 percent, compared with 62 to 63 percent for the original ungated design, and kept attack success at or below 7 percent even when 30 percent of the observations about a target were adversarial. That is the real test for any assistant marketed as having persistent memory: whether a bad actor can quietly rewrite what it remembers.
The authors' own caveat is the interesting part. Once adversarial evidence outpaces honest evidence, the same quorum mechanism that protects memory starts amplifying the attack instead, so 'resistant to misinformation' still depends on who shows up first.