← MCP

Knowledge Graph Memory ServerKnowledge Graph Memory Server

Model Context Protocol

记忆Memory 本地Local 读写Read and write verified lifecycle: active 核验Checked 2026-10-23

用本地知识图谱做跨会话持久记忆的 reference server。它记忆落盘、跨会话共享，但没有访问控制。

A reference server implementing cross-session persistent memory with a local knowledge graph — memory is written to disk, shared across sessions, with no access control.

适合与不适合When to use it, when not to

让 Agent 跨会话记住用户偏好、项目背景等长期信息。

不适合存放需要审计或追溯的场景（图谱会原地修改）， 也不适合存放高敏感信息——记忆文件无加密、无独立访问控制，任何能读该文件的进程都能拿到全部记忆。

Scenarios needing cross-session memory that survives restarts and is queryable as a graph.

Not a fit for anything with access-control requirements — memory is written to disk with no access control.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网仓库在 GitHub，大陆直连波动（同日实测 200/000 混合）；npm / PyPI 安装源实测稳定可达。Hosted on GitHub — direct reachability from mainland fluctuates (200/000 mixed on the same day); npm / PyPI install routes measured stable.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

11 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 本身不调用模型，是被客户端调用的接口实现。

设计与 Claude 的持久记忆场景绑定（README 明言「lets Claude remember information about the user across chats」）。
It calls no model itself; it is an interface implementation called by clients.

Its design is tied to Claude's persistent-memory scenario (the README states plainly "lets Claude remember information about the user across chats").
运行位置Where it runs: 本地 Node.js 进程，通过 stdio 通信。

数据持久化在本地 JSONL 文件，生命周期独立于客户端进程。
A local Node.js process communicating over stdio.

Data persists in a local JSONL file, with a lifecycle independent of the client process.
本地文件Local files: 只访问自己的记忆文件（默认在 memory 目录下），不提供任意文件读写工具。

存储位置可通过参数指定，实际可访问范围取决于该路径的权限。
Accesses only its own memory file (by default under the memory directory); it provides no arbitrary file read/write tools.

The storage location can be set via a parameter; the actual accessible range depends on the permissions of that path.
关机后的任务Tasks after shutdown: server 进程随客户端生死，

但记忆数据持久存在磁盘上，重启后仍在。

2026-09-03 修复了图变更序列化以避免并发写竞争。
The server process lives and dies with the client,

but memory data persists on disk and survives restarts.

A 2026-09-03 fix serialized graph mutations to avoid concurrent-write races.
工具与扩展Tools & extensions: 提供 9 个工具：

create_entities、create_relations、add_observations、delete_entities、delete_observations、delete_relations、read_graph、search_nodes、open_nodes。

另暴露一个 resource：memory://knowledge-graph。

含 3 个删除工具（delete_entities / delete_observations / delete_relations），输出为结构化 JSON，可直接消费。
It provides 9 tools:

create_entities, create_relations, add_observations, delete_entities, delete_observations, delete_relations, read_graph, search_nodes, open_nodes.

It also exposes one resource: memory://knowledge-graph.

Including 3 delete tools (delete_entities / delete_observations / delete_relations); output is structured JSON, ready for direct consumption.
上下文与记忆Context & memory: 这是本 server 的核心能力：跨会话持久记忆。

用知识图谱结构存储实体、关系与观察，支持检索（search_nodes）与按名读取（open_nodes）。
This is the server's core capability: cross-session persistent memory.

It stores entities, relations and observations in a knowledge-graph structure, supporting retrieval (search_nodes) and read-by-name (open_nodes).
权限与限制Permissions & limits: 无独立审批机制。

记忆图谱是全局共享状态，任何写入都会影响后续所有会话。

提供 3 个删除工具，可删除实体、观察与关系。

2026-09-03 增加了 search_nodes 查询长度限制与去重修复。
No separate approval mechanism.

The memory graph is global shared state; every write affects all subsequent sessions.

It provides 3 delete tools that can remove entities, observations and relations.

On 2026-09-03 a search_nodes query-length limit and a deduplication fix were added.
适合什么任务Best for: 让 Agent 跨会话记住用户偏好、项目背景等长期信息。

不适合存放需要审计或追溯的场景（图谱会原地修改）。
Lets an Agent remember long-term information such as user preferences and project context across sessions.

Not a fit for scenarios needing audit or traceability (the graph is mutated in place).
传输方式Transport: 仅 stdio。本地 Node.js 进程，通过标准输入输出通信（npx 启动）。

官方 README（2026-10-08 全文核验）没有远程传输的官方启动方式。
stdio only. A local Node.js process communicating over standard input/output (launched via npx).

The official README (full-text verification 2026-10-08) has no official launch method for a remote transport.
认证机制Authentication: 无独立认证层。stdio 形态继承客户端用户权限。

记忆文件本身无加密，任何能读该文件的进程都能获取全部记忆内容。
No separate authentication layer. In stdio form it inherits the client user's permissions.

The memory file itself is unencrypted — any process that can read that file can obtain all memory content.
权限范围Permission scope: 作用域最窄的一类：只读写自己的记忆文件，不提供任意文件访问工具。

权限风险不在文件范围，而在内容敏感度——

记忆里可能有用户偏好、项目信息等隐私，且文件无加密、无独立访问控制。
The narrowest-scope class: it reads and writes only its own memory file and provides no arbitrary file-access tools.

The permission risk lies not in file scope but in content sensitivity —

the memory may hold user preferences, project information and other privacy, and the file is unencrypted with no independent access control.

头号误解Top misconception

以为记忆只存在于当前会话，实际是磁盘持久化的全局状态Assuming memories live only in the current session — they are disk-persisted global state
以为删除操作需要审批，实际没有，误删会影响后续所有会话Assuming deletion needs approval — it does not, and an accidental delete affects every later session
以为它是简单的 key-value 存储，实际是知识图谱结构Assuming it is a plain key-value store — it is a knowledge-graph structure

未知项清单Known unknowns

记忆文件的默认位置与命名
是否支持 Streamable HTTP（仓库未声明）
存储后端是否可替换（是否只支持 JSONL）
并发写的最终一致性保证程度

The default location and naming of the memory file
Whether Streamable HTTP is supported (not declared in the repo)
Whether the storage backend is replaceable (is JSONL the only supported form)
The degree of eventual-consistency guarantees for concurrent writes

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-08。

类型Type	名称Name	链接Link
docs	MCP · Memory ServerMCP · Memory Server	https://github.com/modelcontextprotocol/servers/tree/main/src/memory
changelog	MCP · Memory 目录变更历史MCP · Memory directory change history	https://github.com/modelcontextprotocol/servers/commits/main/src/memory
changelog	MCP · Servers ReleasesMCP · Servers Releases	https://github.com/modelcontextprotocol/servers/releases
repo	MCP · Servers 仓库 README（生产环境免责声明）MCP · Servers repo README (production-use disclaimer)	https://github.com/modelcontextprotocol/servers

实测记录Measured runs

本档未在真实客户端逐项跑通；工具清单与权限来自官方 README。

This site has not completed measured runs. The test protocol is at protocols/mcp/_protocol.md.

变更记录说明Change-log notes

本对象无独立 changelog 页，变更记录在目录 commit 历史中：

日期	内容
2026-09-03	序列化图变更，避免并发写竞争
2026-09-03	单批次内跳过重复实体与关系
2026-09-03	限制 `search_nodes` 查询长度

三条修复集中在同一天（09-03），说明这一版对稳定性做了集中处理。特别是「并发写竞争」——说明多会话同时读写记忆曾有数据损坏风险。

如果你在多会话环境用它，这三条值得读原文。

This object has no standalone changelog page; change records are in the directory's commit history:

Date	Content
2026-09-03	Serialize graph mutations to avoid concurrent-write races
2026-09-03	Skip duplicate entities and relations within one batch
2026-09-03	Limit `search_nodes` query length

All three fixes landed on the same day (09-03), showing that release concentrated on stability. Especially the "concurrent-write race" — it means simultaneous multi-session reads/writes of memory once risked data corruption.

If you use it in a multi-session environment, these three are worth reading at the source.

数据结构Data structures

概念	说明	示例
Entity	图谱节点，有唯一名称、类型、观察列表	`{name: "X", entityType: "person", observations: [...]}`
Relation	实体间的有向连接	`{from: "A", to: "B", relationType: "works_at"}`
Observation	挂在实体上的原子事实	建议一条观察只存一个事实

关系用主动语态存储，如 works_at 而非 works at。这是文档里的明确约定。

Concept	Description	Example
Entity	a graph node with a unique name, type and observation list	`{name: "X", entityType: "person", observations: [...]}`
Relation	a directed connection between entities	`{from: "A", to: "B", relationType: "works_at"}`
Observation	an atomic fact attached to an entity	one fact per observation is recommended

Relations are stored in active voice, e.g. works_at rather than "works at". This is an explicit convention in the docs.

工具清单与读写属性Tool inventory and read/write attributes

工具	属性	说明
`read_graph`	读	读取整张图
`search_nodes`	读	按条件检索节点
`open_nodes`	读	按名读取特定节点
`create_entities`	写	新建实体
`create_relations`	写	新建关系
`add_observations`	写	给实体追加观察
`delete_entities`	删除	删除实体及其关联
`delete_observations`	删除	删除观察
`delete_relations`	删除	删除关系

6 读 3 写 3 删除。删除操作有三个工具，分别用于节点、观察和关系。

Tool	Attribute	Description
`read_graph`	read	read the whole graph
`search_nodes`	read	search nodes by condition
`open_nodes`	read	read specific nodes by name
`create_entities`	write	create entities
`create_relations`	write	create relations
`add_observations`	write	append observations to entities
`delete_entities`	delete	delete entities and their connections
`delete_observations`	delete	delete observations
`delete_relations`	delete	delete relations

6 read, 3 write, 3 delete. Three deletion tools cover entities, observations and relations.

权限风险的特点What is distinctive about its permission risk

和其他 MCP server 不同，memory 的风险不在文件范围，在内容：

维度	评价
文件访问范围	最窄 — 只碰自己的记忆文件
写入影响	全局 — 影响后续所有会话
内容敏感度	高 — 可能存有用户偏好、项目信息
访问控制	无 — 文件无加密，无独立鉴权

任何能读到那个 JSONL 文件的进程，都能拿到全部记忆。

如果你用它存敏感信息，这一点必须考虑。

Unlike other MCP servers, memory's risk is not in file scope but in content:

Dimension	Assessment
File-access scope	narrowest — touches only its own memory file
Write impact	global — affects all subsequent sessions
Content sensitivity	high — may hold user preferences, project information
Access control	none — the file is unencrypted, no independent auth

Any process that can read that JSONL file obtains the entire memory.

If you store sensitive information with it, this point must be considered.

与 filesystem 的对比Comparison with filesystem

维度	filesystem	memory
访问范围	目录白名单（可很宽）	仅自己的记忆文件（很窄）
删除能力	不提供	提供 3 个删除工具
持久化	无状态	图谱持久化到磁盘
跨会话	不支持	核心能力

取舍很清楚：memory 用「能力受限但内容敏感」换取「跨会话能力」。

Dimension	filesystem	memory
Access scope	directory allowlist (can be wide)	only its own memory file (very narrow)
Delete capability	not provided	provides 3 delete tools
Persistence	stateless	graph persisted to disk
Cross-session	not supported	the core capability

The trade-off is clear: memory trades "limited capability but sensitive content" for "cross-session capability".