How does dual-path parametric memory personalize language models across sessions?

A single recurrent state lets early evidence fade

Long-term personalization asks a language model to carry a user's preferences, background, and constraints from earlier sessions into later answers. One way to avoid re-reading the whole history at inference time is parametric memory: encode the interaction history into model parameters or adapters so the history itself need not sit in the context window. The paper separates two existing routes and names a weakness in each. Context compilation methods such as Doc-to-LoRA and SHINE map individual contexts to LoRA adapters but leave cross-session integration unspecified. State maintenance methods such as RPMem and Metis integrate sessions sequentially, which the authors argue can attenuate earlier evidence.

The paper makes the second concern concrete rather than rhetorical. For a recurrent memory whose state after segment t is an update of the previous state using segment representation q_t, suppose two histories differ at an earlier segment i and share everything after it. If each later update contracts that difference by at most a factor κ below 1, the difference at the final step is bounded by κ raised to the power T−i times the original difference. The older the contribution, the more the bound shrinks. This is a statement about a sufficient condition on the update, not a measurement of any particular deployed system, but it motivates the design that follows.

Evidence preserves, Delta orders, fusion combines

DPPM encodes each segment of the interaction history with a frozen compiler and works on the resulting segment representations. Its Evidence path pools those representations directly. A learned projection and a coordinate-wise temporal softmax assign each segment a weight per latent coordinate, and the weighted sum is rescaled to a calibrated magnitude. Because every segment connects straight to the pooled output, an early segment does not have to survive a chain of updates to influence the final memory. The paper's own bound for this route depends on history length T but not on segment position, so earlier and later segments share the same lower bound on pooling weight. The authors are explicit that this mitigates age-related attenuation rather than eliminating forgetting, since the bound still falls as T grows.

Pooling alone has a structural blind spot: it produces the same result when fixed segment representations are reordered, so it cannot distinguish two opposite update orders that require different answers. The Delta path supplies that ordering. It processes the same representations chronologically, derives a key and a write gate from each segment, reads a prediction of the current segment from an associative state, and writes back a gated residual. The state is therefore corrected toward what the current segment actually says, and key similarity controls how much one write interferes with another key's readout. A learned fusion then mixes the calibrated Evidence and Delta outputs coordinate by coordinate, and a frozen decoder turns the fused vector into history-conditioned LoRA factors. Only the lightweight memory modules are trained; the compiler, decoder, shared adapter, and backbone stay frozen.

Gains concentrate on forgetting and attribution

On Qwen3-4B, DPPM scores 54.22% on PersonaMem-v2 and 86.79% on PrefEval, above the evaluated baselines across multiple backbones. The category breakdown is more informative than the headline averages. Against RPMem, DPPM gains 21.68 points on the Forget category, which tests whether responses respect requests to stop using prior information; 17.57 points on Sensitive, which tests restraint with private details; and 15.34 points on Therapy, which tests personalization from individual background. The Others subset, which tests whether information about another person is wrongly attributed to the user, rises from 13.68% to 50.23%.

The authors read this pattern as better attribution and more appropriate revision of memory use, which fits their motivation of preserving evidence while revising its applicability. That is their interpretation, and the category design supports it: the largest movements are on questions about when remembered information should and should not be applied, not on every information type. The same figure also shows small changes elsewhere, including a 1.05-point decrease on the Stereotypical category, so the improvement is not uniform across all seven PersonaMem-v2 information types. On PrefEval, gains appear in all four held-out topics, from 1.90 points on Technology to 6.50 on Education.

Latent retention separates the two paths

To test the attenuation argument directly, the paper measures latent sensitivity as distractor interactions accumulate. It freezes checkpoints, places one preference among 32 fixed distractors after a shared neutral opening segment, and keeps each history at 34 segments. For paired preference values A and B, sensitivity is the norm of the difference between the two consolidated representations divided by the norm of the difference between the two preference-segment representations. Each profile is normalized by its age-zero value and averaged over 24 profiles, so the reported numbers are retention ratios, not accuracy scores.

At age 32, DPPM retains 0.970 on PersonaMem-v2 checkpoints and 0.667 on PrefEval checkpoints. RPMem retains 0.268 and 0.116, and a Delta-only control retains 0.367 and 0.300, while an Evidence-only control stays at 1.000. The ordering is the point: the direct pooling route is the least age-sensitive, the sequential route loses the most, and the fused model sits between them but far above the recurrent baseline. This is a controlled latent measurement on fixed checkpoints, not a claim that answer accuracy stays constant as histories grow, and the PrefEval-trained ratio of 0.667 shows that fusion does not fully preserve the Evidence-only level.

Answer latency stays low while memory stays compact

The efficiency results concern the answer stage after memory preparation, with compilation and retrieval excluded. DPPM improves accuracy over RPMem by 8.50 percentage points at an answer-forward latency of 46.14 ms, only 2.43 ms higher. It is also more accurate than RAG and Full Context while computing answers roughly 3.1 times and 62.7 times faster, respectively.

For longer histories, the paper selects one question at the official 128K history length and evaluates both official history versions. Full Context inputs grow from 33,143 to 139,319 tokens, pushing peak GPU memory from 10,203.91 to 18,245.97 MiB. DPPM stays at 7,762.32 MiB, matching RPMem and using 57.46% less memory than Full Context on the longer history. These figures come from a single length-selected question on one A800 with batch size 1 and KV cache disabled, so they describe a compact answering footprint rather than a full serving benchmark.

Training objective matters less than the dual-path design

The paper also checks whether DPPM depends on a particular training signal. Holding the architecture fixed, plain cross-entropy scores 54.22% on PersonaMem-v2 and 86.79% on PrefEval, averaging 70.51%. Answer forward KL and short on-policy distillation with 16-token rollouts both fall below cross-entropy on both datasets. The 64-token on-policy distillation configuration reaches 54.56% and 87.04%, the highest two-dataset average at 70.80%, while one-step GRPO reaches the highest PersonaMem-v2 score at 54.80% but drops PrefEval to 85.06%, pulling its average below cross-entropy.

The authors conclude that appropriately configured on-policy feedback can complement the dual-path architecture while plain cross-entropy remains competitive with simpler training. The practical reading is that the architecture, not the objective, carries most of the gain: no variant moves either benchmark by more than about one point relative to cross-entropy, whereas the gap to RPMem on PersonaMem-v2 is 8.50 points. The 64-token distillation result also uses a revised teacher prompt, so it is not a clean comparison of objective alone.

What the evidence covers and what it does not

The evaluation spans PersonaMem-v2 and PrefEval across three backbones of different architectures and sizes from 2B to 7B parameters, and the reported headline numbers use Qwen3-4B. Both benchmarks test preference following and implicit personalization from interaction history; neither establishes how DPPM behaves for other kinds of long-term memory, such as factual recall over months of conversation, or for multi-user settings where histories interleave. The latency and memory figures come from a single-GPU, batch-size-1 answer stage with compilation and retrieval excluded, so they do not describe end-to-end system cost.

The paper's own framing is also modest about retention: direct accumulation reduces age-related attenuation but does not remove the dependence on history length, and the PrefEval-trained retention ratio of 0.667 at age 32 shows the fused model still loses some sensitivity. Within those bounds, the most decision-relevant comparison is between the two paths rather than between training recipes. Evidence-only preserves latent sensitivity perfectly but cannot represent order; Delta-only captures order but decays fastest; DPPM's fused adapters land closest to the retention of the pooling route while still gaining the attribution and revision improvements that the category results attribute to ordered updates.

Sources used in this answer