Vector + graph memory backend: observe/remember/search/recall tools against a local memoria server, fusing HNSW semantic recall, FTS5 keyword search, and knowledge-graph signals via RRF ranking, with automatic turn-end persistence and per-agent namespace isolation.
Install
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:jiayan-xu/dsh-memoria
GitHub-sourced plugins run build scripts on your machine at install time. Only install sources you trust, and pin a commit (github:owner/repo#sha).
README
memoria 记忆后端插件:把 memoria(向量 + 图记忆层)接入 DeepSeek Harness (dsh),让 dsh agent 会话可以记住和回忆。
- 4 个工具:
memoria_observe/memoria_remember/memoria_search/memoria_recall - 自动写入:每轮对话结束自动
observe沉淀;用户肯定反馈(不错/很好/good/赞…)自动remember(importance=5) - 配置热重载:改
~/.dsh/settings.yaml的memoria:section 免重启生效 - 命名空间隔离:所有读写强制落在配置的 namespace(默认
dsh-test),不碰其他业务数据
使用前提
⚠️ 本插件是 memoria 的前端,不是记忆存储本身。使用前需要:
- 自建并运行 memoria 服务(默认连接
http://127.0.0.1:9003):memoria 是独立的开源项目(jiayan-xu/memoria,Rust 向量 + 图记忆服务)。没有该服务时工具会报连接失败。 - 准备 badge token:memoria 工具调用要求
register_agent签发的 badge token(admin key 不能直连),见下文「获取 memoria badge token」。以环境变量MEMORIA_AGENT_KEY提供(不落盘)。
装完发现工具报 Connection refused / HTTP 200 but auth failed,先检查这两项。
安装
前置:@deepseek-ai/dsh CLI 与 pnpm。
# 安装(自动挂载,无需手动改 profile)——两种源任选
dsh plugin --profile web add github:jiayan-xu/dsh-memoria
# 或 npm 源:
dsh plugin --profile web add @jhp830901/dsh-memoria
# 启动 dsh web 即可使用
dsh web
配置
插件默认连 http://127.0.0.1:9003(本地 memoria)。可配置项:
| 项 | 默认 | 说明 |
|---|---|---|
baseURL |
http://127.0.0.1:9003 |
memoria 服务地址 |
agentId |
dsh-memoria |
memoria agent 身份(需已注册) |
apiKeyEnv |
MEMORIA_AGENT_KEY |
环境变量名,值 = memoria badge token |
namespace |
dsh-test |
记忆命名空间(隔离边界) |
autoWrite |
true |
自动写入开关 |
推荐用 settings section 覆盖(热重载):
# ~/.dsh/settings.yaml
memoria:
baseURL: http://127.0.0.1:9003
namespace: my-namespace
autoWrite: true
agentKey 不要落盘,用环境变量:
export MEMORIA_AGENT_KEY=<badge-token>
获取 memoria badge token
memoria 的工具调用要求 X-Agent-Key 为 register_agent 签发的 badge token(admin key 不能直连)。未注册时先注册:
# 用 admin key 注册 agent(示例),返回 badge 有效期 1 年
curl -X POST http://127.0.0.1:9003/mcp -H 'Content-Type: application/json' \
-H 'X-Agent-Id: dsh-memoria' -H 'X-Agent-Key: <ADMIN_KEY>' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"register_agent","arguments":{"agent_id":"dsh-memoria","display_name":"dsh-memoria plugin","namespace":"dsh-test"}}}'
使用
用户:记住一个重要事实:我的仓库密码是 42
用户:我之前记住的仓库密码是什么? → 命中回忆
验收用例
- 写入:
memoria_remember {content, namespace}→{"status":"remembered","id":...} - 召回:
memoria_search {query, namespace}→ 语义 + 关键词多信号命中 - 隔离:其他 namespace 检索本插件数据 → 授权拒绝/空结果
开发
npm install -D typescript @types/node --legacy-peer-deps
bash scripts/build.sh # 产出 lib/(Windows 下依赖 junction,见 scripts/build.sh 注释)
License
BSD-3-Clause
Links
More in this category
LoserFox/distill★ 15
Automatic conversation distillation: background subagent reflection + skill create/update.
omdsh-dev/dsh-mnemon★ 11
Deep Mnemon integration: local three-tier memory (Runtime Memory, retrievable Documents, supervised Memory Spaces).
modusensus/dsh-mneme★ 8
Cross-session memory: SQLite with a human-editable Markdown mirror, background consolidation (dedup, merge, conflict resolution), and six memory tools.
nowledge-co/nowledge-mem-deepseek-harness★ 5
One memory layer for every AI tool and agent: Context Bundle injection, prompt-time recall, MCP tools, and turn-end DSH thread capture.
Jesse-njx/dsh-memory★ 2
Cited memory over DSH's lossless session log: distilled facts carry `(sessionId, eventRange)` citations that expand back to the exact original log excerpt.
PerryLink/dsh-memento★ 1
Bounded, layered, approval-gated, auditable cross-session memory: a typed `ctx.memory` seam with a zero-dependency SQLite provider, a `memory` tool, and frozen snapshot injection; every write passes the approval gate and stays reconstructable from the session log.