LLM agents' memories become outdated, rather than hallucinating—and it's nearly impossible to predict this

When people talk about memory issues in LLM agents, they usually mean write-time hallucinations: the distiller model invents a fact that wasn’t present in the conversation. This is a well-known issue; there are benchmarks for it (HaluMem), and write-time gates are designed to address it. But there’s a second problem that almost no one addresses: a fact is recorded perfectly and still turns out to be false. You moved. You changed jobs. You sold your laptop. Nothing was compromised, nothing was hallucinated—the world simply moved on, but the memory didn’t. The creator of the ANAMNESIS project encountered this firsthand: he told his assistant that he’d switched to a MacBook; the consolidation module recorded the new fact thirteen seconds before the next question; the retrieval module pulled it into context—yet the answer still said “ThinkPad.” No component had broken down. It’s just that none of the components was tasked with noticing that the old fact was outdated. ANAMNESIS is a local memory system for long conversations with LLMs, built around three mechanisms. GATE checks whether a fact was supported by the source at the time of recording (not a new idea—ProMem did this earlier). PREDICT attempts to forecast whether a specific piece of clean memory will become corrupted within the next N turns. HEAL performs repairs triggered by the forecast. PREDICT is the only truly new component, and it’s the one that isn’t working yet. The predictor was trained end-to-end in three seconds on a CPU, at no cost whatsoever. The forward AUROC on a strictly delayed split is 0.665. This means: given two memory chips, one will fail and the other won’t—the model ranks them correctly about two out of three times. Better than random chance, but far from useful. Logistic regression on the same features yielded 0.6647. Gradient boosting outperformed the simplest baseline by 0.0007—a tie. But the most valuable insight wasn’t found in the metrics, but when testing the trained model against the actual database schema. The dominant feature accounted for 76.8% of the model’s gains—and it was the dataset annotation, which is a hard-coded constant in production. The model found a shortcut, as models tend to do when production features remain unfilled. The calibration score looked flawless (ECE 0.00012) until we built the reliability diagram—all predictions fell into a single cluster. A model that says the same thing about everything is trivially calibrated and carries no information. Three conclusions that extend beyond the scope of this project. First: an obvious feature may be incorrect. The instinct is to use embedding distance and look for outliers. But the MINJA attack shows that poisoned memories, implanted through ordinary conversation, are intermixed with benign ones in the embedding space, because the attack works through blending. A detector based on this geometry is blind to the very attack it is supposed to catch. Therefore, the features here are behavioral and structural, with no embedding distance at all. Second: a scalar metric can be fooled by a degenerate model. The calibration number was excellent and meaningless. Any scalar can hide a model that has learned to say one thing—the diagram behind it cannot. Third: exporting data to a table and examining it is more effective than reading the code. Two actual bugs in the project were found this way, rather than by reading the source code: shortcut learning and a string mismatch that silently reset three features across 14,823 lines. The project is a working document for my final year, not a finished paper. The results are “scored pending”; automatic HEAL is part of the v2 workstream. But the honesty of the report is more valuable than a number that might look better. Source: Hacker News — ANAMNESIS

Source: Hacker News - Newest: ""AI" "LLM""