Hy-Memory/memhub/docs/ARCHITECTURE.md

MemHub Architecture

Updated: 2026-10-21.

Hy Memory 是 SDK / 引擎；MemHub 是产品与发布名。one-memory 仍可能出现在 兼容路径、模块名和服务标识中。

---

Architecture Principles

1. 所有对 memory 的读写操作都交给 SDK，包括存储 backend；backend 选择与路径一律通过 SDK config / env 完成（如 vector / graph / cache / history / coding / raw）。MemHub 不另起一套 memory 存储引擎。
2. MemHub 内对 memory 的原子操作也一律经 SDK 完成。 SDK 若缺少所需 client 能力，先补 SDK 接口，再在产品层调用；禁止在 Core 内直接改 VDB/Graph/SDK 私有库，或再维护一份 memory 正文权威。
3. MemHub 层不存储 memory 实体与引擎侧关系（正文、supersede 权威、向量/图副本等）。MemHub 的主责是 管理面：harness（来源授权与同步）、LLM model profile、异步写入队列、config 开关、Core/Runtime 状态。
   • 可以且应当持久化 SDK 提供的 memory_id（及组织用元数据），作为外键引用。
   • 禁止再存一份 memory 正文台账、自判 supersede、把 SDK 当可丢缓存。

直接约束：

• 原始对话进入 SDK raw 层；MemHub 不以 evidence 表保存原文权威。
• 列表、纠正、删除和召回通过 授权 + SDK API 完成；需要挂接处只存 memory_id。
• Matter 尚未形成已发布产品合同；当前全部 matter* 占位表和 API 入口删除。

---

Product and Runtime Shape

MemHub is a local-first memory product.

Responsibilities

AI hosts / harnesses
  ├─ install / hooks / MCP / Skill / pairing
  ├─ authorized recall (Host Bridge → Core → SDK search)
  └─ future higher-level organization requires a new explicit contract
             │
             ▼
      Python Core / FastAPI   ← 管理面
      ├─ harness grants & sync desired-state
      ├─ model profiles & secrets boundary
      ├─ async write queue (schedule / drain / leases)
      ├─ config switches & Core/Runtime health
      ├─ Hy Memory bridge (only SDK client entry for memory I/O)
      └─ local HTTP API
          ▲        ▲         ▲
          │        │         └─ Host Bridge / recall broker
       Web UI    CLI / Runtime Control
                    ▲
                 Electron

Memory entities / raw / vectors / graph / coding stores
  └─ owned exclusively by Hy Memory SDK (config/env under One Memory Home)

There is one managed Runtime and one MemHub data home per installation. CLI and Electron are clients and lifecycle interfaces to that Runtime; they do not own separate memory engines. The Web UI is a client of Core and does not access host files or SDK private stores directly.

Workspace Ownership

Path	Responsibility
apps/core	FastAPI 管理面：harness、model、队列、config、Runtime 协作，并经 bridge 调用 SDK。
apps/web	React / TypeScript product UI and API client.
apps/desktop	Electron shell; attaches to or starts the managed Runtime.
apps/cli	Bootstrap, install, upgrade, service lifecycle, diagnostics.
packages/runtime-control	Runtime discovery/lifecycle, install transactions, service management.
packages/host-bridge	Loopback recall broker and host-facing MCP/Skill/Hook/plugin assets. Not a second memory engine.
packages/contracts	Shared TypeScript API contracts.
packages/operations	Shared typed operation flows.
scripts	Dev/privacy/test/packaging gates; not runtime product code.

Python: uv + 3.12. TypeScript: pnpm + Node 24. See root package.json, apps/core/pyproject.toml, ../README.md, ../CONTRIBUTING.md.

Data and Processing

Source authorization and synchronization（管理面 · 仍属 MemHub）

Source root authorization and Harness synchronization remain distinct:

• SQLite（或后续等价管理库）source_grants：授权根、身份、授予/撤销。Grant alone does not turn sync on.
• Core-owned <data_dir>/config.json：per-source sync_enabled 等 desired state。
• Core is the only writer of config.json（temp + fsync + atomic replace）.
• Web/Desktop/CLI/Runtime call Core APIs; they do not keep a second durable copy.

同步/导入把宿主对话写入 SDK（含 raw 层）。队列描述待执行的 SDK 调用，完成态以 SDK 可验证结果和可选 memory_id 回执为准。

暂停同步是管理面开关；已进入 SDK 的 memory/raw 生命周期由 SDK API 管理， 不能通过 MemHub 影子表单独改变其可见性。

Periodic polling / fs watch live in Core. UI does not schedule scans.

Memory processing

Write path

authorized harness / user action
  -> (optional) durable job row in MemHub queue  // 调度与租约，不是 memory 实体
  -> sedimentation / model / source gates         // 管理面门禁
  -> HyMemoryBridge
  -> Hy Memory SDK write (raw / chat / coding / …)
  -> SDK stores (vector, graph, cache, history, coding, …) via SDK config/env
  -> MemHub may persist memory_id only for explicit management receipts / UI keys
  -> search / recall / UI via SDK read APIs (+ MemHub authz)

• 原子 memory 变更（create / update / delete / supersede 等）：只通过 SDK。缺接口则扩展 SDK client。
• 异步队列：只保留 memory_write_jobs，其承载 FIFO / event-time、公平调度、租约、暂停、重试与 SDK 回执；一个 job 对应一次 SDK add()。旧 memory_tasks 与 task receipt 二级台账删除。
• SDK write_outcomes 只回答某个 request 是否已由 SDK 提交及返回哪些 memory_id；它不是队列，也不接管调度。
• 原文/raw 的持久化与失败保留由 SDK raw 合同承担。
• 运维状态直接来自 memory_write_jobs，不另设 memory_tasks 二级执行表。

HY / SDK and MemHub bridge

Hy Memory is the only extraction, memory entity, relationship, and semantic-retrieval engine. Product code enters it through HyMemoryBridge（共享 Runtime、模型配置、密钥边界），不在业务模块里私自 new 一套绕过门禁的 client。

• Storage backends：仅 SDK config/env（Chroma/FAISS/Qdrant、Kuzu/Neo4j、cache/history sqlite|mysql、coding sqlite 等）。
• Queue：pending memory_write_jobs 只表示尚未调用 SDK 的待处理请求，因此没有 SDK raw/memory ID。worker 单次调用 SDK add()；SDK 内部完成 raw 写入与 extraction/reconciliation，成功后 MemHub 仅将返回的 raw ID / memory IDs 记作 job/task outcome。import_raw() 只用于历史数据的 exact migration，不拆分正常写路径。
• Recall：Host Bridge / Core 先做宿主安装凭据与 tool grant，再调 SDK search/get。当前产品没有逐 memory 的 A-harness→B-harness 可见性开关，默认允许跨 harness 搜索；未来若增加来源限制，直接透传 SDK source / include_sources / exclude_sources，不在 MemHub 建 memory ACL。
• Context wire：cutover marker 提交前只使用 legacy Context v1；提交后 Host Bridge 只使用 SDK-native Context v2。两种 schema 原样校验、原样返回，不把 SDK item 补造成 v1 的 revision、evidence_ids、paths 或 adoption，已选择的 v2 路径失败时也不回退 legacy 权威。
• Ask wire：marker 前保留 Governance Ask v1；marker 后 Web / Host Bridge 只调用 SDK Ask v2。v2 的引用只含 SDK 原生 memory_id/content/layer/status/source/gmt_created，生成过程中重复 get 校验同一 SDK 快照，不补造 revision、Evidence 或 scopes，SDK 失败不回退 v1。
• Matter：未形成已发布产品合同，不迁移占位数据；所有 matter* 表在升级清理时物理删除，fresh install 不创建，Runtime 不注册 Matter API。
• 旧用户迁移：由 App 在可见的启动 loading 阶段统一执行；此时 Core 尚未对外提供产品读写。启动流程创建/校验完整 bundle、续跑 ledger、完成 SDK backfill 与 reconciliation，满足全部门禁后才提交 marker 并进入主界面。普通用户不需要另跑 CLI，也不按旧 App 内 SDK 版本分叉迁移。
• Matter 若未来重新开发，需以新的 SDK-native 合同重新设计，不继承当前占位 schema。

Pinned SDK 版本见发布锁文件 / MEMORY.md（随发布更新）。

Permission Boundaries（管理面 · 仍独立）

These permissions answer different questions and must not be collapsed:

1. Source grant: may Core read this selected host data root?
2. Harness sync: should Core continue discovering/importing from this source into the SDK write path?
3. Model-content permission: may eligible content be sent to the configured model provider (for SDK extraction etc.)?
4. Host recall grant: may this external host invoke the granted recall tools? This is not a per-memory cross-harness ACL.

Host installation / recall authorization cross an external trust boundary; preview, fingerprint, scope, credential, rollback protections remain.

The product must not depend on screen capture, Accessibility/AXTree, OCR, keyboard/mouse, clipboard, or background audio collection.

Storage

存什么	谁拥有	怎么配
Memory 实体、raw、向量、图、coding、SDK cache/history	Hy Memory SDK	SDK config / env；目录落在 One Memory Home 下由 Runtime 约定
harness grants、sync desired-state、model profiles、队列/租约、Core 收据、必要的 memory_id 引用	MemHub 管理库（当前多为 one-memory.sqlite3）	Core 管理；不存 memory 正文

MemHub SQLite（或后继）是管理面权威库；HY/SDK stores 是 memory 数据平面 权威。两者职责不同，MemHub 不镜像 memory 实体。

Fresh install 直接 bootstrap 最新管理面 schema，不执行历史 Evidence / Governance migration 链。旧用户启动迁移可以临时读取旧表，但 cutover、raw 引用回填和 bundle restore 验证全部成功后，必须在 Runtime 对外服务前 DROP 旧 Evidence/Governance/retrieval/session-shadow、全部 Matter 占位表以及迁移专用 alias/ledger。

不要为 MemHub 管理面引入隐式 MySQL 双写。SDK 侧 MEMORY_CACHE_BACKEND=mysql 等属于 SDK 合同，与 MemHub 管理库无关。

config.json：desired user settings，不是 runtime health 的唯一真相；健康态从收据与进程状态重建。

管理面可持久化 source_grants、model profile、队列/租约、call_events 审计元数据（不存对话或记忆正文）和 Runtime 收据。

Matters / call_events / session inbox（口径）

能力	当前口径
Matters	未发布占位功能，现已删除表与 Runtime API；未来按新合同重新设计。
call_events	本机调用元数据审计（非 memory 实体）；可留在管理面；展示记忆时用 id 调 SDK。
session inbox / lease / change_outbox	已删除的历史管道。未来若需要宿主增量，必须定义管理事件或 SDK 可订阅合同，不得靠 MemHub 影子 memory 表推送正文。

Runtime and Interfaces

The repository has one Runtime implementation, one default data home, and one OS service identity per installation. macOS App and npm CLI share Runtime artifact contract, ~/.one-memory, receipts, ports, credentials, Core API, single-writer lock, stable install root, and service declaration.

Runtime process shape

The long-running Runtime is a Node process running runtime.cjs serve. It owns:

• one Python Core child;
• one recall Broker / Host Gateway sidecar;
• Runtime/Core/Broker/token/lock receipts under the data home;
• Core restart supervision while Runtime remains alive.

Electron UI and npm CLI discover Runtime via receipts and verify loopback identity before attaching. Core’s data-directory lock prevents a second Core writer; it does not replace a single OS startup owner.

macOS App path

A production App under /Applications performs a managed installation before attaching its UI:

1. Copy the bundled Runtime into ~/Library/Application Support/One Memory/{versions,current}.
2. Install the stable ~/.local/bin/memhub link.
3. Write an App-owned service receipt under ~/.one-memory.
4. Install or migrate ~/Library/LaunchAgents/com.tencent.onememory.runtime.plist.
5. Wait for Runtime, Core, and Broker to become verifiably ready.
6. Attach Electron as a client of that Runtime.

The LaunchAgent starts current/bin/node current/runtime/runtime.cjs serve directly. It does not start Electron and does not open a window:

macOS user login
  -> com.tencent.onememory.runtime LaunchAgent
  -> managed Runtime
  -> Python Core + Broker

Electron App -> attach/control only

The old SMAppService.mainAppService registration is read only for migration. Closing all windows or quitting Electron does not stop the managed Runtime. Login-start preference controls runtime-service.enabled / launchd KeepAlive.PathState. App deletion removes the enable marker; user Memory data and stable Runtime files remain until explicit uninstall/delete-data.

npm Bootstrap / CLI path

Lightweight npm Bootstrap CLI plus independently acquired Runtime. Bootstrap does not ship Python Core, Web UI, Host Bridge, models, or full Runtime. memhub install --yes resolves/verifies release, installs Runtime, links CLI, registers user service, starts, waits for Core readiness (rollback-capable).

On macOS, same user LaunchAgent path; CLI-managed service uses KeepAlive.PathState + enable marker. On Linux, systemd user service with restart-on-failure. npm uninstall removes Bootstrap only; memhub uninstall removes service/Runtime/links, keeps data by default; delete-data is separate.

Current lifecycle matrix

Scenario	Startup owner	After user login	Full UI quit	App deletion	Process-level recovery
macOS App-owned	LaunchAgent	Runtime/Core/Broker without Electron	Runtime continues	Marker removed; Runtime stops	launchd ↔ Runtime ↔ Core
macOS CLI-owned	Same LaunchAgent	Runtime/Core/Broker without Electron	No Electron dependency	GUI deletion no effect	launchd ↔ Runtime ↔ Core
Linux CLI-owned	systemd user	Runtime with user manager	No UI dependency	N/A	systemd ↔ Runtime ↔ Core

Service ownership

macOS LaunchAgent or approved managed helper
Linux systemd user service
              -> managed Runtime
              -> Core + Broker

Electron App -------- attach/control only
npm CLI ------------- attach/control only

Owner receipt is app or cli. Host Bridge exposes only loopback / host-scoped capabilities; it does not persist a general-purpose Core bearer as an unrestricted proxy.

Current Support and Release Boundary

Support is declared per immutable GitHub Release manifest:

• Linux x64/arm64 Runtime on native GHA runners.
• macOS arm64/x64 App/Runtime: Developer ID, notarization, native acceptance.
• Windows is not supported.
• hy-memory==2.0.5.dev3 is the current development-preview SDK dependency; the release lock and ../MEMORY.md are authoritative when the pin changes.

See INSTALLATION.md.

Engineering Rules

• Architecture Principles 不可被「方便测试 / 赶发布」绕过。
• Memory I/O 与 backend 只经 SDK；缺能力先补 SDK。
• MemHub 可存 memory_id 引用与管理面状态；不存 memory 实体副本。
• Source grant、model egress、host recall scope 保持独立。
• 不要把「管理库查不到行」当成 SDK 侧已删除，除非已显式调 SDK 删除并确认。
• Do not treat unavailable source data as deletion（宿主文件消失 ≠ 自动删 SDK memory，除非产品显式定义并调 SDK）。
• Do not let graph layout imply semantic relationships.
• Run the smallest relevant tests for ordinary development; full/release/soak only when requested.
• Never remove a safety boundary merely to make a gate green（授权/模型/召回边界）。

Test entry points: ../CONTRIBUTING.md.

Upgrade Compatibility

Released installations require the offline, resumable compatibility protocol in contracts/sdk-data-plane-migration-v1.md. The App runs it during the startup loading phase before Core exposes product reads or writes. A cutover requires a complete bundle, resumable backfill, reconciliation, and verified cleanup; a SQLite-only backup or replaying legacy rows through extraction is not sufficient.