Agent memory needs a bouncer: WikiTrace and FGSE check lessons before trusting them
Memory makes agents worse more often than vendors admit. The KAIST team behind WikiTrace (https://arxiv.org/abs/2611.14320, code at https://github.com/yukyeongleee/wikitrace) names it memory-induced degradation: the agent fails a task it would have solved with no memory at all, because it followed a wrong lesson or misread a right one. The root cause is simple. Lessons are distilled from runs labeled only pass or fail, then handed to the agent stripped of the run they came from, so there is nothing to check them against.
WikiTrace keeps the receipts. Each lesson is stored as a short insight linked to a summary of why its source run succeeded or failed, and to the raw trajectory. On a new task the agent starts from the insights and, when one is vague, conflicts with another, or came from a failed run, it drills down through the links with terminal tools before acting. Across eight suites built from ALFWorld, ScienceWorld, PlanCraft, AppWorld and tau-bench Retail, WikiTrace averages 90.2 against 79.5 with no memory and beats the same insights without links by 3.2 points. The number that matters more: regression, failing tasks the no-memory agent solves, drops to 4.5%, versus 6.5% for insights alone and 9.7% for simply rerunning the no-memory agent. It is the only method in the paper that matches or beats no memory on every suite.
Falsification-Gated Self-Evolution (https://arxiv.org/abs/2611.14326, Tang et al.) attacks the same problem at write time instead of read time. Every candidate lesson becomes a hypothesis with a precondition, a behavior change, an expected effect and a verifier. FGSE applies it in a temporary state, runs transfer, falsification and archived-regression probes, then commits, refines or rejects it. Reported results: 39.8 success on WebArena, 78.8 average completion on AppWorld, 0.724 and 0.487 Pass^1 on tau-Retail and tau-Airline, with falsification failures at 5.0%, regression damage at 1.3% and harmful committed updates at 3.6%.
One checks lessons when reading, the other when writing. Both treat memory like code going into production: no evidence, no merge. That is the right instinct. A memory file an agent writes for itself is the least-reviewed code in your stack, and two papers in one day just showed how cheap it is to put a gate in front of it.
← Back to all articles
WikiTrace keeps the receipts. Each lesson is stored as a short insight linked to a summary of why its source run succeeded or failed, and to the raw trajectory. On a new task the agent starts from the insights and, when one is vague, conflicts with another, or came from a failed run, it drills down through the links with terminal tools before acting. Across eight suites built from ALFWorld, ScienceWorld, PlanCraft, AppWorld and tau-bench Retail, WikiTrace averages 90.2 against 79.5 with no memory and beats the same insights without links by 3.2 points. The number that matters more: regression, failing tasks the no-memory agent solves, drops to 4.5%, versus 6.5% for insights alone and 9.7% for simply rerunning the no-memory agent. It is the only method in the paper that matches or beats no memory on every suite.
Falsification-Gated Self-Evolution (https://arxiv.org/abs/2611.14326, Tang et al.) attacks the same problem at write time instead of read time. Every candidate lesson becomes a hypothesis with a precondition, a behavior change, an expected effect and a verifier. FGSE applies it in a temporary state, runs transfer, falsification and archived-regression probes, then commits, refines or rejects it. Reported results: 39.8 success on WebArena, 78.8 average completion on AppWorld, 0.724 and 0.487 Pass^1 on tau-Retail and tau-Airline, with falsification failures at 5.0%, regression damage at 1.3% and harmful committed updates at 3.6%.
One checks lessons when reading, the other when writing. Both treat memory like code going into production: no evidence, no merge. That is the right instinct. A memory file an agent writes for itself is the least-reviewed code in your stack, and two papers in one day just showed how cheap it is to put a gate in front of it.
Comments