面向智能体的蒸馏式长期记忆：按语义召回、按证据门控写入，每种智能体模式对应一个 kura。原生 DeepSeek Harness 插件，无需向量数据库、嵌入模型或框架。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-09-04
许可证
MIT
01它能帮你完成什么？
Give each DeepSeek Harness agent preset its own persistent memory store
A separate kura store per agent mode, recalled by meaning, so a preset change switches what the agent remembers
DSH users running multiple agent presets who want mode-scoped long-term memory
Write agent knowledge without poisoning the store
Memories distilled and poured only after an evidence gate (class, candidate, judge), not blind accumulation
Builders who want durable, auditable agent memory instead of unchecked assertions
02如何接入 DeepSeek Harness？
前置条件
- Python 3.11+ — no third-party dependencies (`[dev]` adds only pytest)
- Node.js 20+ — required only for the DSH plugin
- An OpenAI-compatible model endpoint answering `POST <url>/chat/completions`
安装步骤
- 01
$ git clone https://github.com/lna-lab/distill-kura && cd distill-kura
- 02
pip install -e . # or just run: python3 -m distill_kura.cli
- 03
cp kura.example.toml kura.toml # edit: one model endpoint is enough to start
- 04
kura init main --path ~/kura/main # create an empty store
- 05
kura serve # http://127.0.0.1:8085
验证接入成功
- curl -s -X POST localhost:8085/recall -H 'content-type: application/json' \
03DSH 适配与能力边界
Loads as a native DeepSeek Harness plugin through `@deepseek-ai/dsh-tools` `defineTool`, exposing `kura_*` tools bound to a preset's memory store.
Semantic recall (recognition)
a question, or a named memory→picked memory slugs plus their [[link]] neighbourhood as context
Evidence-gated writing
a journal of agent events→new memories poured only after class → candidate → gate → judge
Per-mode memory stores
an agent preset / mode→a bound kura store; a preset change switches the active memory
Ships as DSH plugin, MCP server, HTTP service, Python library
any host→the same engine behind a DSH plugin, MCP, HTTP (:8085) or import
04适合谁？何时不该用？
适合
- DSH users running multiple agent presets who want mode-scoped long-term memory
- Builders who want durable, auditable agent memory instead of unchecked assertions
不适合
- Requires an OpenAI-compatible `POST <url>/chat/completions` model endpoint; the API key is read from an environment variable you name, never stored in config.
05兼容性、维护与安全提示
- The DSH plugin needs Node.js 20+; the core engine is Python 3.11+ with no third-party dependencies.
- Requires an OpenAI-compatible `POST <url>/chat/completions` model endpoint; the API key is read from an environment variable you name, never stored in config.
- Standard library only — no vector database, no embeddings, no framework; recognition uses a small local model over the whole index in one prompt.
Active repo (master), MIT licensed, no tagged release as of 2026-09-03.
06常见问题
如何把 distill-kura 接入 DeepSeek Harness？
它作为原生 DSH 插件通过 `@deepseek-ai/dsh-tools` 的 `defineTool` 加载，暴露绑定到预设仓库的 `kura_*` 工具。在预设中指定一个仓库，切换预设即可切换整个记忆体。
没有嵌入模型，召回是怎么做到的？
整个索引（每条记忆一行）放进一次提示，由小模型识别与问题相关的记忆；命中具名记忆则是确定性命中。README 称约 500 条记忆的索引约 6k token。
记忆写入安全吗？
写入受证据门控：日志先分类、生成候选、通过 GATE，再经撰写、草稿、评判，最后才倒入；不会把朴素断言直接记成事实。
运行需要什么环境？
Python 3.11+ 且无第三方依赖；仅 DSH 插件需要 Node.js 20+；以及任意兼容 OpenAI 的模型端点（`POST <url>/chat/completions`）。
07相关的 DSH 工作流
weknora
作者 tencent
开源的 LLM 知识库平台：把原始文档转化为可检索的 RAG 知识库、自主推理智能体和自动维护的 Wiki 系统，支持多租户部署。
honcho
作者 plastic-labs
用于构建有状态 AI 智能体的记忆库，支持跨会话连续学习与上下文工程。
mirage
作者 strukto-ai
全球首个面向 AI 智能体的统一虚拟文件系统。
reme
作者 agentscope-ai
ReMe：面向 AI Agent 的记忆管理套件——帮 Agent「记住我、优化我」，支持 RAG 与长期记忆。
08数据与来源
Ships as a [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) plugin,
DSH switches **persona and tools** by agent preset. distill-kura switches **memory** by
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-09-04。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。