Research map / Large language models / Agents, code and math
ParamMem: Augmenting Language Agents with Parametric Reflective Memory
Tianjun Yao, Yongqiang Chen, Yujia Zheng, Pan Li, Zhiqiang Shen, Kun Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ParamMem parametric memory to Reflexion (ParamAgent) | Llama-3.1-8B agent · HumanEval | Pass@1: 76.22 -> 82.93 | Improved |
| add ParamMem to Reflexion (ParamAgent) | Llama-3.1-8B agent · 2WikiMultiHopQA | accuracy: 78.67 -> 88.67 | Improved |
| add ParamMem to DoT-bank cross-sample memory (ParamAgent-plus) | Llama-3.1-8B agent · MATH | accuracy: 73.02 -> 75.45 | Improved |
| use diversity-driven ParamMem instead of accuracy-driven Retroformer module | Llama-3.1-8B agent · MBPP | Pass@1: 42.82 -> 67.00 | Improved |
| use Retroformer rather than ParamMem on math | Mistral-7B-v0.2 agent · MATH | accuracy: 43.53 -> 28.06 | Hurt |
| train ParamMem on data generated by Llama-3.1-8B itself (self-improvement) | ParamAgent · HumanEval | Pass@1: 73.17 -> 78.05 | Improved |
| train ParamMem on self-generated data, plus cross-sample memory | ParamAgent-plus · HotpotQA | accuracy: 72.00 -> 83.33 | Improved |
| add Llama-3.1-8B-based ParamMem to a Qwen3-Next-80B agent | ParamAgent-plus · LiveCodeBench | Pass@1: 62.00 -> 63.33 | Improved |
| use Qwen3-Next-30B ParamMem instead of Llama-3.1-8B ParamMem for coding (ablation) | ParamAgent-plus on Qwen3-Next-80B · LiveCodeBench | Pass@1: 63.33 -> 68.00 | Improved |
| use Qwen3-Next-30B ParamMem instead of Llama-3.1-8B on QA (ablation) | ParamAgent-plus on Qwen3-Next-80B · HotpotQA | accuracy: 85.00 -> 83.67 | Hurt |
| train ParamMem on 500 clustered samples instead of 8000 (ablation) | ParamAgent · HumanEval | Pass@1: 82.93 -> 81.71 | No clear effect |
| add cross-sample memory to ParamAgent trained on 500 samples (ablation) | ParamAgent-plus · HumanEval | Pass@1: 81.71 -> 86.59 | Improved |
Closest papers by meaning
- M$^\star$: Every Task Deserves Its Own Memory Harness (NeurIPS 2026 · similarity 0.89)
- Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (NeurIPS 2026 · similarity 0.89)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.89)
- AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.89)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.89)
- PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
