RPMem: Learning Long-Term Recurrent Parametric Memory Across Sessions for LLM Agents
Fanyu Zhao, Ruike Cao, Liang Dong, Fugen Yao, Jian Xu, Guanjun Jiang, Han Zhang, Yifei Zhao, et al.
cs.CL · cs.AI
Quick verdict
Claims backbone-reusable recurrent parametric memory; strong, well-ablated gains on PERMA but modest on other benchmarks; multiple-choice-only evaluation.
Summary
The paper introduces RPMem, a two-stage parametric memory architecture for LLM agents that decouples memory from the serving backbone. Stage one trains a hypernetwork memory compiler—a frozen context encoder, a Perceiver resampler, and a model-specific LoRA decoder—to map each session into a fixed-shape latent memory via one forward pass, using a fixed-reference forward-KL objective that matches context-conditioned next-token distributions on a top-32 support plus tail mass. Stage two trains a lightweight recurrent gate (roughly 0.5M parameters) that selectively integrates each incoming session memory with the accumulated state; at inference the consolidated memory is decoded into backbone-specific LoRA parameters, and backbone replacement requires only retraining the decoder. On PERMA with Qwen3-8B, RPMem reaches 85.52% core average, exceeding Metis-9B by 5.32 and Full Context by 12.98 points; ablations credit the learned gate with +31.31 points over factor averaging. Across five backbones, decoder transfer raises the cross-model average from 75.97% to 89.64%, while updates cost 0.043 s at 64 sessions within a fixed 36 MiB per-user memory. Gains narrow elsewhere: PersonaMem-v2 Overall is 40.22% against a 57.90% oracle, and LightMem beats RPMem on Pref Eval at 10- and 70-turn horizons. Evaluation is multiple-choice answer selection, main tables report no significance tests, and PERMA uses only ten users.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- RPMem is the first lifecycle-independent parametric memory architecture: the accumulated latent memory, session encoder, and gate persist unchanged across serving-backbone replacements.
- Fixed-reference forward-KL compilation is decisively better than alternatives, reaching 85.52% PERMA core average versus 61.21% for Compiler SFT.
- The task-trained coordinate-wise recurrent gate is essential, beating the best fixed consolidation rule (factor averaging, 54.21%) by 31.31 pp.
- Memory writes are forward-only with near-constant cost: 0.043 s per session and 36 MiB fixed per-user memory at 64 sessions, with at most 3.81 pp general-capability loss.
- Model-specific decoding transfers one memory representation across five backbones, raising the cross-model PERMA average from 75.97% to 89.64%.
Strengths
Genuinely new capability with paired transfer evidence
Backbone replacement is tested as a matched protocol (Appendix F): keeping encoder and gate frozen and retraining only the decoder improves all 28 target-backbone x setting combinations and lifts the cross-model PERMA average from 75.97% to 89.64%.
Ablations isolate both stages convincingly
Table 2/C.1 varies the compilation objective (Fixed FKL 85.52% vs Compiler SFT 61.21%, Top-K CE 77.66%); Table 2/C.2/C.3 varies consolidation (learned gate 85.52% vs best fixed rule 54.21%, tuned EMA 53.51%, task-trained scalar gate 53.39% under matched supervision).
Deployment costs are measured, not asserted
Exact recurrence enables real mechanistic analysis
Unusually transparent protocol checks
Weaknesses
Headline gains are benchmark-concentrated
Multiple-choice-only evaluation
Thin statistical reporting for main claims
Uneven baseline adaptation
Per-scenario consolidation and Qwen-centric data
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RPMem (arXiv:2609.23466v2, Fudan University + Qwen/Alibaba) targets parametric memory for long-running LLM agents and makes three claims: (1) a two-stage architecture that compiles sessions into a fixed-shape latent memory by forward computation and integrates them with a task-trained recurrent gate; (2) lifecycle independence - the first parametric memory that survives serving-backbone replacement by decoding a model-independent latent memory into backbone-specific LoRA parameters; (3) validation across three benchmarks (PERMA, PersonaMem-v2, Pref Eval) and five backbones. The headline - 85.52% on PERMA with Qwen3-8B, +5.32 pp over Metis-9B and +12.98 pp over Full Context (Table 1) - is real and internally consistent, but it is also the most favorable cell in the paper, as detailed below.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is RPMem and what problem does it solve?
RPMem is a two-stage parametric memory system for LLM agents. Instead of storing conversation history as text to retrieve at query time, it compiles each session into a fixed-size latent memory via one forward pass, merges it into an evolving memory with a learned recurrent gate, and converts that memory into LoRA adapter parameters for the current model. This removes history tokens from the prompt and lets accumulated memory survive when the serving model is replaced (only the small decoder is retrained).
How much better is RPMem than baselines?
On PERMA with Qwen3-8B it reaches 85.52% (core average), beating the strongest parametric baseline Metis-9B by 5.32 points and Full Context by 12.98 points. The advantage is smaller elsewhere: PersonaMem-v2 Overall is 40.22% (vs 31.10% Full Context and a 57.90% oracle), and on Pref Eval the text-based LightMem system is ahead at 10- and 70-turn horizons.
Does RPMem still work after the underlying LLM is swapped?
Yes - that is its core claim. The latent memory, session encoder, and gate are model-independent; only a backbone-specific LoRA decoder is retrained. In the paper's paired transfer experiments across four target backbones (including Ministral-3-8B and Qwen3.5-35B-A3B), decoder transfer improved all 28 backbone-setting combinations, raising the average from 75.97% to 89.64%.
Is the code and data available, and what does it cost to run?
The paper states an implementation is available (the URL is garbled in the extracted text). Full compiler training is expensive - about 338 GPU-hours on 8 GPUs - but refitting the consolidation gate for one PERMA fold takes about 13 minutes on a single A800-80GB, and all hyperparameters, data sources, and prompts are documented in the appendices.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-3 and C.1-I.1 appear only as extracted axis/label text; visual claims (survival curves, violin distributions) could not be verified from images.
- Several passages were omitted in extraction (parts of section 3.2, Appendix B.2, D.3-D.4, F.1); some protocol details are inferred from surrounding text.
- The code URL is truncated to 'available at code.' in the extraction, so repository existence was not verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.