- Rating
- 3
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper studies long-term memory for multimodal agents in the setting where a general fact about an entity coexists with exceptions that hold under conditions visible in a query image (a "conditional conflict"). It argues that prior systems lose the applicable record either at write time (overwrite, expiry, or merge) or at read time (similarity ranking), and proposes ConeMem, which stores each entity as a subtree in the Lorentz model of hyperbolic space: the general fact sits near the origin and exceptions branch outward inside entailment cones whose apertures are fixed by position alone. A trained observer (frozen SigLIP 2 encoders, one cross-attention block, and a projection) maps the image region selected by an exception's condition text to a point in the same space, and applicability is read as a cone margin, so the choice among coexisting records requires no language-model call. A local write procedure (reinforce, branch, or pending, with containment repair and bounded replay) keeps the hierarchy consistent, while a frozen text-only reasoner performs extraction, same/contrary labeling, entity naming, and answering. Proposition 1 shows that nesting consumes an additive radial budget under negative curvature and a multiplicative one in Euclidean space. Under a leave-one-benchmark-out protocol on ATM-Bench, Mem-Gallery, MemEye, and H2HMem with two reader families, ConeMem improves macro-average judge accuracy by 13 to 17 points over the strongest published baseline at the lowest answer-stage cost of the compared systems, comes within 2 to 3 points of exhaustive pairwise verification, and is accompanied by extensive ablations, three Euclidean controls, a curvature sweep, controlled depth and branching experiments, an oracle decomposition, and sensitivity, scale, and label-noise studies.
Critical strengths
S1. Well-posed problem with a crisp diagnosis. The conditional-conflict failure mode is formalized as three requirements, connected to the classical specificity preference of nonmonotonic reasoning, and the predicted error signature is verified on annotator-labeled questions. This is a good motivation.
S2. The design is coherent, and each component is tied to a requirement. The "no language-model call for applicability" property is real, and the cost accounting is complete.
S3. The evaluation is broad and mostly careful: four benchmarks, two reader families plus a third text-only reasoner, leave-one-benchmark-out with two alternative protocols.
S4. The "geometry is load-bearing" claim is tested rather than asserted: three Euclidean controls that keep the tree, observer, and training. a curvature sweep, controlled depth and branching subtrees, and cone diagnostics showing that apertures become level-dependent only after write training.
Critical weaknesses
W1. Hyperparameter selection appears to use the evaluation benchmarks. In every row of Table 12 the default value is the argmax of the full average over the four evaluated benchmarks. Leave-one-benchmark-out protects the observer's weights but says nothing about these constants, and Section 4.1 only states that the same values are used on the four benchmarks. Were the constants chosen on the training benchmarks' questions, on a validation split, or on the reported test questions? This could be a serious problem. Please explain.
Questions for authors
See Weakness please. Does hyperparameter selection use the evaluation benchmarks? If so the results of the proposed method might be untrustworthy. Please explain my only question honestly and thoroughly. Thanks in advance.