DSH | dsh-memory | 结合 cc 的 auto-memory 与 codex 的 session 整理的长期记忆 (附 LoCoMo 测评结果)

[hr98w]

非官方项目，由社区成员独立开发和维护。 / Unofficial community project.

项目 / Repository: https://github.com/hr98w/dsh-memory
测评报告 / Interactive report: https://hr98w.github.io/dsh-memory/

中文

dsh-memory，给 DeepSeek Harness 加上跨 Session 的长期记忆。实现尽量简单：记忆就是本机的 Markdown，不需要向量数据库或独立记忆服务。

核心思路：Claude Code × Codex

灵感来源	借鉴的思路	dsh-memory 的实现
Claude Code Auto Memory	Markdown 索引与渐进式披露	提供 Global 记忆与当前 Workspace 的 MEMORY.md，Agent 按需读取详细文件
Codex Session 记忆整理	从历史对话提炼可复用信息	用户手动选择历史 Session，由独立整理 Agent 更新 Global 与来源 Workspace 记忆

Global 保存跨项目偏好，Workspace 保存项目事实和约定。日常 Agent 通过 memory_update 写入记忆；新 Session 按需读取。模型提出修改，Host 负责校验、版本检查和写入。

插件接入 DSH 的提示词、工具和 Settings 页面，支持中英文 UI。整理模型直接从 DSH 已配置的模型中选择。

LoCoMo-10 测评

也跑了一次完整测评，想看看这些 Markdown 能否帮助回答历史对话里的问题。流程参考了 OpenViking 的 Claude Code LoCoMo 设置，不作跨项目分数比较。

使用 DeepSeek V4 Flash，排除 Category 5 后共 1,540 题，用 LLM Judge 对照标准答案评分：

模式	问答前做什么	正确率
Baseline	不提供历史或记忆	2.60%
Memory	按顺序导入历史，Agent 保存记忆	73.05%
Memory + Consolidate	独立导入历史形成记忆，再整理 Session	77.60%

这说明在这套设置下，记忆能帮助回答历史问题，但不是与完整上下文或其他记忆系统的对比。整理并非每组都更好；两种记忆模式独立形成初始记忆，不能把分数差全归因于整理。

受成本限制，每种模式只保留一组完整结果，含开发过程中的重试和重跑；其中 Memory 的 conv-41 在观察到低分后重跑并替换，可能存在选择偏差。Ingest、QA、Judge 都有模型随机性。逐题回答、分类分数和成本已放在在线测评报告里。

English

dsh-memory adds long-term memory across Sessions to DeepSeek Harness. The implementation stays simple: memory is stored as local Markdown, with no vector database or separate memory service required.

Core idea: Claude Code × Codex

Inspiration	Idea	Implementation in dsh-memory
Claude Code Auto Memory	Markdown index and progressive disclosure	Provide Global memory and the current Workspace's MEMORY.md; read details on demand
Codex Session consolidation	Extract reusable knowledge from past conversations	Manually selected Sessions are processed by a separate Agent that updates Global and source-Workspace memory

Global holds cross-project preferences; Workspace memory holds project facts and conventions. Agents write through memory_update and recall relevant files in later Sessions. The Host validates proposals, checks revisions, and commits changes.

The plugin integrates with DSH prompts, tools, and Settings, with Chinese/English UI and a consolidation model selector using existing DSH model configuration.

LoCoMo-10 evaluation

I also ran a complete evaluation to see whether these Markdown memories help answer questions about past conversations. The workflow follows the approach of OpenViking's Claude Code LoCoMo setup, without comparing scores across projects.

Using DeepSeek V4 Flash, the evaluation covers 1,540 questions after excluding Category 5, with an LLM Judge scoring answers against the reference answers:

Mode	Before QA	Accuracy
Baseline	No history or memory provided	2.60%
Memory	Ingest history in order; the Agent saves memories	73.05%
Memory + Consolidate	Independently ingest history to form memories, then consolidate the Sessions	77.60%

These results suggest memory helps with historical QA in this setup, not that it beats full context or other memory systems. Consolidation did not improve every conversation, and the independently built memories prevent a clean causal comparison.

Due to cost constraints, each mode retains only one complete result set, including development retries and reruns. Memory's conv-41 was rerun and replaced after a low score, introducing possible selection bias. Ingestion, QA, and judging all involve LLM variability. Per-question answers, category scores, and costs are available in the interactive evaluation report; further limitations are documented in the experiment retrospective (Chinese).

安装 / Install

dsh plugin --profile web add @hr98w/dsh-memory

安装后完全重启 dsh web。 / Fully restart dsh web after installation.

目前请让一个进程负责同一 DSH_HOME 的记忆写入，跨进程并发保护尚未实现。
Use one memory-writing process per DSH_HOME; cross-process write protection is not implemented yet.

截图 / Screenshot