← News

Agents & MCPSources: arXiv · 18 Nov 2026

RPMem compiles agent sessions into reusable parametric memory

The paper's authors present a two-stage memory architecture for long-running LLM agents that survives backbone replacement and keeps update costs near-constant as histories grow.

In plain words

AI agents that work over many sessions need a way to remember past interactions. RPMem turns each session into a compact internal memory that is not tied to any specific model, then adapts it to whatever model is in use. The authors say this lets memory persist and be reused even when the underlying model is swapped out.

What was announced

The authors argue that text-based memory, which retrieves and reconstructs past interactions at every query, becomes increasingly dependent on retrieval quality and contextual reasoning as histories grow. Parametric memory instead encodes experience directly into model computation, but they note existing approaches offer limited support for cross-session evolution and are coupled to a specific backbone, restricting reuse after model replacement.

RPMem works in two stages: it compiles each session into a model-independent latent memory through forward computation, then selectively integrates it with retained memory using a task-trained recurrent gate. The consolidated memory is mapped to backbone-specific low-rank adaptation (LoRA) parameters, so the encoding capability transfers when the backbone changes. In the paper's evaluation across three long-term memory benchmarks and five backbones, the authors report near-constant update cost and memory footprint, and with Qwen3-8B on the PERMA benchmark they report 85.52%, exceeding the strongest parametric and text-based baselines by 5.32 and 12.98 percentage points respectively. Ablations, they say, support the complementary roles of session compilation and cross-session consolidation, and dynamics analyses suggest the gate learns task-specific integration strategies. Code is available on GitHub.

What it does not show

  • Evidence is limited to the paper's abstract; the full methodology, experimental setup and results have not been reviewed here.
  • All performance figures and comparisons are the authors' own reported results and have not been independently confirmed.

For security and IT teams

Scope
The source is an arXiv abstract (v3, replace announcement) covering the paper's summary of RPMem, not the full paper.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.