Files
2026-07-10 18:52:00 +08:00

38 lines
5.9 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# 证据账本
`full` 表示核过实验表、关键消融或 limitation;`context` 表示全文已转文本并做章节定位,但不承担本轮核心结论。
| Date | Paper | Depth | 可支持的判断 | 不能外推的部分 |
| --- | --- | --- | --- | --- |
| 2025-01 | [Zep, 2501.13956](https://arxiv.org/abs/2501.13956) | full | 长对话静态 QA 可用时间图谱大幅缩短上下文和延迟,并提升部分题型 | DMR 已饱和;single-session-assistant 反而下降;不是行动任务 |
| 2025-02 | [A-MEM, 2502.12110](https://arxiv.org/abs/2502.12110) | full | 动态 linked-note 组织在 LoCoMo 上能优于若干早期 memory baseline | 方法、模型和 token 配置并非严格单变量控制 |
| 2025-05 | [Experience-Following, 2505.16067](https://arxiv.org/abs/2505.16067) | full | 经验会影响后续相似任务,错误和错配经验会传播 | 不能证明某种通用写入或删除策略最优 |
| 2025-07 | [MemoryAgentBench, 2507.05257](https://arxiv.org/abs/2507.05257) | full | retrieval、TTL、LRU、forgetting 是不同能力;复杂更新和全局综合仍弱 | 数据集转换和统一 prompt 不能覆盖真实 agent loop |
| 2026-02 | [MemoryArena, 2602.16313](https://arxiv.org/abs/2602.16313) | full | 通用 memory/RAG 在相互依赖的多 session 行动任务上成功率低,额外 memory 延迟更高 | benchmark 本身很难,失败同时含 policy/reasoning 问题 |
| 2026-03 | [Retrieval vs Utilization, 2603.02473](https://arxiv.org/abs/2603.02473) | full | LoCoMo 上 retrieval 方法差异大于 write strategyraw RAG 是强 baseline | 单模型、单 benchmark、固定 top-k=5 |
| 2026-05 | [Cross-Component Interference, 2605.05716](https://arxiv.org/abs/2605.05716) | full | memory 与其他 scaffold 的边际作用可为负,必须做 subset ablation | 只有 HotpotQA/GSM8Ktool 组件含协议因素 |
| 2026-05 | [EvoMemBench, 2605.18421](https://arxiv.org/abs/2605.18421) | full | 15 种方法间不存在跨能力通用最优 memory form | 统一 benchmark 仍不能消除所有实现质量差异 |
| 2026-05 | [DeMem, 2605.10870](https://arxiv.org/abs/2605.10870) | full | 决策充分状态比描述相似性更适合行动记忆;LoCoMo 描述相似性相关很弱 | 主要是自定义决策任务和理论建模 |
| 2026-05 | [MemQ, 2605.08374](https://arxiv.org/abs/2605.08374) | full | provenance DAG credit assignment 在多步任务有增益,单步接近无增益 | 没有处理删除和长期错误积累 |
| 2026-06 | [AutoMEM, 2606.04315](https://arxiv.org/abs/2606.04315) | full | 跨场景无通用赢家;索引会产生 schema losspassive retrieval 会产生 retrieval loss | 新增 DCI harness 与已有系统的工程成熟度不同 |
| 2026-06 | [MAGE, 2606.06090](https://arxiv.org/abs/2606.06090) | full | MemoryArena 同设置下,active execution-state tree 同时提高 SR 并减少 token | 单团队结果;绝对成功率在 Travel 仍只有 0.1519 |
| 2026-06 | [Infini Memory, 2606.10677](https://arxiv.org/abs/2606.10677) | full | topic document 的 maintenance 对结果贡献明显,memory 维护不应被当后台清理 | 系统组件较多,维护和检索仍依赖 LLM |
| 2026-06 | [HORMA, 2606.11680](https://arxiv.org/abs/2606.11680) | full | organize 与 retrieve 可拆分;高能力 manager 比大 retriever 更关键 | 结论依赖其层级 memory 和任务设置 |
| 2026-06 | [Online Budget Study, 2606.15017](https://arxiv.org/abs/2606.15017) | full | budget-matched 时,vanilla agent 在 WebArena 多模型汇总上可胜 online skill/memory | 只研究 online augmentation,不覆盖离线学习 |
| 2026-06 | [TrustMem, 2606.25161](https://arxiv.org/abs/2606.25161) | full | transition-level verifier 和排名训练可降低 omission/corruption/hallucination | verifier 仍是模型;训练、judge 和跨域泛化需要独立核验 |
| 2026-06 | [Supersede, 2606.27472](https://arxiv.org/abs/2606.27472) | full | 当前值维护是独立瓶颈,更强模型和更大 memory 不足;训练信号能推动它 | 训练只有 Qwen2.5-3B 单次运行,16.7% 远非解决 |
| 2026-06 | [MemDelta, 2606.29914](https://arxiv.org/abs/2606.29914) | full | embedding、model behavior 和 write cost 足以翻转架构结论 | Mem0 只在 88 题、2/6 题型上做 matched comparison |
| 2026-06 | [ATMem, 2606.31612](https://arxiv.org/abs/2606.31612) | full | GUI agent 的主动任务状态和选择性调用能提高结果 | 128xH20 训练;更难 DataScope 仍接近失败 |
| 2026-06 | [GovMem, 2607.02579](https://arxiv.org/abs/2607.02579) | full | 自动写入必须检查独立证据、反例、scope 和 provenance;真实轨迹暴露 false promotion | 外部 133 条是 stress slice,安全自动覆盖率为 0 |
| 2026-07 | [A-TMA, 2607.01935](https://arxiv.org/abs/2607.01935) | full | current/historical/transition 角色能缓解 ghost memory | 收益 host-dependentLoCoMo 部分指标不一致 |
| 2026-07 | [AgenticSTS, 2607.02255](https://arxiv.org/abs/2607.02255) | full | typed memory contract 让技能/episode 可消融和复现 | 3/10 到 6/10 不显著;缺 same-code accumulating-context baseline |
| 2026-07 | [FARMA, 2607.05029](https://arxiv.org/abs/2607.05029) | full | 伪造推理记录是不同于事实投毒的写入攻击面 | 50-trial 自定义攻击;0% ASR 防御不等于开放环境稳健 |
## 全文池
本轮全文池共 32 篇:
`2501.13956`, `2502.12110`, `2505.16067`, `2507.02259`, `2507.05257`, `2507.07957`, `2511.20857`, `2512.18746`, `2601.01885`, `2602.16313`, `2602.19320`, `2603.02473`, `2603.07670`, `2603.15666`, `2605.05716`, `2605.08374`, `2605.10870`, `2605.14421`, `2605.18421`, `2606.04315`, `2606.06090`, `2606.10677`, `2606.11680`, `2606.15017`, `2606.25161`, `2606.27472`, `2606.29914`, `2606.31612`, `2607.01935`, `2607.02255`, `2607.02579`, `2607.05029`
其中没有进入上表核心论证的 9 篇暂标为 `context`,不把“下载过 PDF”写成“已得到结论”。