feat(copaw-plugin): MemOS Cloud memory backend for CoPaw (demo) by fancyboi999 · Pull Request #1447 · MemTensor/MemOS

fancyboi999 · 2026-04-09T09:05:03Z

背景

Issue #1440 提到希望 MemOS 支持 CoPaw 平台。CoPaw 官方已有基于 ReMeLight 的本地记忆系统，架构上也预留了可插拔的 memory backend 接口（BaseMemoryManager + memory_manager_backend 配置）。

这个 PR 是一次集成尝试（demo 级别），验证了 MemOS Cloud 作为 CoPaw memory backend 的可行性。

做了什么

在 apps/memos-copaw-plugin/ 下新建了一个 CoPaw 插件，包含 4 个文件：

plugin.json — CoPaw 插件清单
plugin.py — 插件入口，通过 api.register_memory_manager("memos", MemOSMemoryManager) 注册 backend
memos_memory_manager.py — 继承 ReMeLightMemoryManager，只覆写 memory_search 走 MemOS Cloud API，其他本地能力（compaction、token 计数、context 管理）全部复用 ReMeLight
memos_client.py — MemOS Cloud HTTP 客户端（/search/memory + /add/message）

用户安装后，在 CoPaw agent 配置里设 memory_manager_backend: "memos" + MEMOS_API_KEY 环境变量即可启用。

当前集成范围

老实说，目前只接了 MemOS 最基础的一层：

能力	状态
`/search/memory` 云端记忆召回	✅ 通过覆写 `memory_search` 实现
偏好召回（preference_detail_list）	✅ 搜索结果已包含
云端不可达时 fallback 到本地 ReMeLight	✅
`/add/message` 对话自动上传	❌ CoPaw 没有 per-turn hook，未实现
工具记忆、技能记忆	❌ 未接入
知识库检索、多 agent 隔离	⚠️ 配置字段预留了，但没实际使用
Recall Filter（LLM 二次过滤）	❌ 未实现

核心缺口是 add 流程 — CoPaw 的 BaseMemoryManager 没有 per-turn 回调，对话内容不会自动上传到 MemOS Cloud。目前只能通过外部手段（MemOS 控制台、API）写入记忆。

本地验证

语法检查 + 单元级导入验证通过
实际启动 CoPaw，插件正常加载，memory_search 走 MemOS Cloud 成功召回记忆
不相关对话不受记忆影响，上下文无污染

配套 PR

CoPaw 侧需要一个小改动来支持插件注册 memory backend：agentscope-ai/CoPaw — feat/memos-memory-backend 分支（PluginApi.register_memory_manager + _resolve_memory_class 扩展）。

想讨论的方向

add 流程怎么做？ CoPaw 需要 post-turn hook 还是有其他方式？
MemOS 的高级功能（工具记忆、技能记忆、Recall Filter）值不值得在 CoPaw 侧接入？ 如果值得，接口怎么设计比较好？
这个插件应该放在 MemOS 仓库还是独立仓库？

cc @tangg555 @CaralHsi @xyxy 想听听你们的想法。

Closes #1440

Implements CoPaw's BaseMemoryManager interface, delegating long-term memory search/add to MemOS Cloud API while handling context compaction locally. Follows the same pattern as existing OpenClaw plugins. Closes MemTensor#1440

…re parity Replace the standalone BaseMemoryManager implementation with inheritance from ReMeLightMemoryManager. This ensures all local operations (token-aware compaction, check_context, compact_tool_result, get_in_memory_memory) use CoPaw's battle-tested ReMeLight code. Only memory_search and summary_memory are overridden for MemOS Cloud.

summary_memory only fires during context compaction, not every turn. Putting MemOS Cloud add logic there was both unreliable (too infrequent) and architecturally wrong (mixing local summary with cloud upload). v1 scope: MemOS Cloud provides recall (memory_search) only. All local operations (compaction, summary, persistence) stay with ReMeLight unchanged. Cloud-side add requires CoPaw to expose a post-turn hook first.

Copilot

Pull request overview

Adds a demo CoPaw plugin that registers a MemOS Cloud–backed MemoryManager so CoPaw agents can recall long-term memory via the MemOS Cloud /search/memory API while reusing ReMeLight’s local memory features and providing a local fallback path.

Changes:

Introduce a CoPaw plugin manifest + entrypoint that registers the "memos" memory backend.
Add MemOSMemoryManager (extends ReMeLightMemoryManager) overriding memory_search to call MemOS Cloud with fallback to local search.
Add an aiohttp-based async MemOS Cloud client (search + add endpoints, plus a lightweight ping).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 6 comments.

File	Description
apps/memos-copaw-plugin/plugin.json	Defines CoPaw plugin metadata and dependency on `aiohttp`.
apps/memos-copaw-plugin/plugin.py	Plugin entrypoint; registers the MemOS memory backend and a startup hook.
apps/memos-copaw-plugin/memos_memory_manager.py	MemoryManager implementation integrating MemOS Cloud recall with local ReMeLight fallback.
apps/memos-copaw-plugin/memos_client.py	Async HTTP client for MemOS Cloud API endpoints used by the memory manager.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-09T09:09:29Z

apps/memos-copaw-plugin/memos_memory_manager.py

+"""MemOS Cloud memory manager for CoPaw agents.
+
+Extends ReMeLightMemoryManager — all local operations (context compaction,
+token counting, tool result truncation, in-memory memory) are delegated
+to the parent class unchanged.  Only two methods are overridden:
+
+  - memory_search  → queries MemOS Cloud instead of local vector index
+  - summary_memory → uploads conversation to MemOS Cloud after local summary
+
+This ensures full compatibility with CoPaw's MemoryCompactionHook and
+force_memory_search auto-recall mechanism.


The module docstring says this class overrides summary_memory to upload conversations to MemOS Cloud, but the implementation does not override summary_memory and the later comment explicitly states cloud "add" is intentionally not implemented. Please update the docstring to match the actual behavior (cloud recall only) to avoid misleading users.

Copilot · 2026-04-09T09:09:30Z

apps/memos-copaw-plugin/memos_memory_manager.py

+from agentscope.message import Msg, TextBlock
+from agentscope.tool import ToolResponse


Msg is imported but never used. This will fail Ruff (F401) in CI; please remove the unused import or use it if intended.

Copilot · 2026-04-09T09:09:30Z

apps/memos-copaw-plugin/memos_memory_manager.py

+            "memory_limit_number": cfg.get("memory_limit_number", 9),
+            "include_preference": cfg.get("include_preference", True),
+            "preference_limit_number": cfg.get("preference_limit_number", 6),
+            "relativity": cfg.get("relativity", 0.45),
+            "timeout": cfg.get("timeout", 8.0),
+            "conversation_id": cfg.get("conversation_id", ""),
+            "knowledgebase_ids": cfg.get("knowledgebase_ids", []),
+            "async_mode": cfg.get("async_mode", True),
+        }


memory_limit_number (and async_mode) are resolved from config/env but never used; memory_search always passes max_results to MemOS instead. Either wire these config values into the request defaults (e.g., use mc["memory_limit_number"] when max_results isn't explicitly provided) or remove the unused config fields to avoid a misleading configuration surface.

Copilot · 2026-04-09T09:09:30Z

apps/memos-copaw-plugin/memos_client.py

+import asyncio
+import logging
+import time
+from typing import Any, Dict, List, Optional


time is imported but never used. This will fail Ruff (F401) in CI; please remove the unused import.

Copilot · 2026-04-09T09:09:31Z

apps/memos-copaw-plugin/memos_client.py

+    async def ping(self) -> bool:
+        """Lightweight connectivity check via a minimal search call."""
+        try:
+            session = await self._ensure_session()
+            async with session.post(
+                f"{self.base_url}/search/memory",
+                json={"user_id": "_ping", "query": "ping"},
+                timeout=aiohttp.ClientTimeout(total=5),
+            ) as resp:
+                return resp.status < 500


ping() treats any HTTP status < 500 as "healthy". If the API key is missing/invalid, MemOS will likely return 401/403, which currently reports as healthy and leads start() to log "connected" even though requests will fail. Consider returning True only for 2xx (or at least treating 401/403 as unhealthy) so connectivity/auth problems are surfaced correctly.

Copilot · 2026-04-09T09:09:31Z

apps/memos-copaw-plugin/plugin.json

+  "name": "MemOS Cloud Memory",
+  "version": "0.1.0",
+  "description": "MemOS Cloud memory backend for CoPaw — provides cloud-based long-term memory via MemOS API (search + add).",
+  "author": "MemTensor",


The plugin description claims "search + add", but this demo implementation does not implement automatic /add/message uploading (per the PR description and memos_memory_manager comment). Please adjust the description to avoid implying that add/sync is supported.

- SIM105: use contextlib.suppress instead of try-except-pass - F401: remove unused imports (Msg, time) - ping(): return True only for 2xx, not all < 500 - plugin.json: remove "add" claim from description - docstring: remove stale summary_memory override reference - wire memory_limit_number config into search request - ruff format fixes

fancyboi999 · 2026-04-09T09:15:17Z

All review feedback addressed in 32a89db:

Stale docstring — removed summary_memory override reference, now accurately describes search-only override
Unused imports — removed Msg and time (F401)
memory_limit_number unused — now wired into search request as max(max_results, config_limit)
ping() too permissive — changed to return True only for 2xx responses
plugin.json claims "search + add" — updated to "cloud-based long-term memory recall via MemOS search API"
ruff format — ran ruff check --fix, 29 auto-fixes applied, 0 remaining

fancyboi999 · 2026-04-10T04:28:48Z

Hermes 的本地插件（memos-local-plugin v1.0.2）已经发了，回来看这个 CoPaw 的 demo，觉得有些地方需要重新想想。

两个框架的记忆体系差得挺远

扒了一圈两边的代码，发现一个容易踩的坑：不能把 Hermes 的模式直接搬到 CoPaw 上。

Hermes 的 MemoryProvider 给了很完整的生命周期钩子 — sync_turn 每轮写入、prefetch 预检索、on_memory_write、on_session_end。MemOS 本地插件的 ingest 管线、混合检索都能直接挂上去。

CoPaw 这边不一样。BaseMemoryManager 的重心在上下文窗口管理：compact_tool_result 截断工具输出、check_context 算 token、compact_memory 做 LLM 摘要压缩。长期记忆检索 memory_search 只是接口里的一个方法，没有 per-turn 回调。

照搬 Hermes adapter 的架构会水土不服。

MemOS 能帮 CoPaw 做什么

memory_search 是匹配度最高的点。MemOS 的混合检索（FTS5 + 向量 + RRF + MMR + 时间衰减）比 ReMeLight 默认的搜索强不少，现在 PR 里继承 ReMeLight 只换掉 memory_search 的思路是对的。

记忆写入可以做，但得找对时机。CoPaw 没有 per-turn hook，不过 compact_memory 是个切入口——上下文压缩时 ReMeLight 会产出结构化摘要（Goals / Progress / Key Decisions），质量不错，可以顺手推到 MemOS。做法是覆写 compact_memory，先 super() 拿到摘要，再异步发到 MemOS，主流程不受影响。

上下文管理（compact、check_context、token 计数）不该碰，这是 ReMeLight 自己的活。文件式持久化（MEMORY.md、每日日志）也不该替代，MemOS 的 SQLite + 向量库是补充。

另一个想法：支持本地模式

现在只走 Cloud API。但 memos-local-plugin 的 Bridge daemon 跑起来已经挺稳了，没理由不让 CoPaw 也用。可以加个 mode 配置：cloud 走现有的 MemOS Cloud，local 走 Bridge daemon，后者能拿到本地混合检索、去重、Web 面板这些。用户按场景选就行。

几个拿不准的

compact_memory 作为写入时机合不合适？CoPaw 侧有没有更好的 hook 我没看到？
本地模式要依赖 Node.js 跑 Bridge daemon，CoPaw 用户群对这个接不接受？
插件继续放 apps/memos-copaw-plugin/ 还是挪到 memos-local-plugin/adapters/copaw/？

cc @tangg555 @CaralHsi @xyxy

fancyboi999 added 3 commits April 9, 2026 16:12

feat(copaw-plugin): add MemOS Cloud memory backend for CoPaw

914ae4e

Implements CoPaw's BaseMemoryManager interface, delegating long-term memory search/add to MemOS Cloud API while handling context compaction locally. Follows the same pattern as existing OpenClaw plugins. Closes MemTensor#1440

Copilot AI review requested due to automatic review settings April 9, 2026 09:05

Copilot started reviewing on behalf of fancyboi999 April 9, 2026 09:05 View session

Copilot AI reviewed Apr 9, 2026

View reviewed changes

style(copaw-plugin): apply ruff format

837a901

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(copaw-plugin): MemOS Cloud memory backend for CoPaw (demo)#1447

feat(copaw-plugin): MemOS Cloud memory backend for CoPaw (demo)#1447
fancyboi999 wants to merge 5 commits intoMemTensor:mainfrom
fancyboi999:feat/copaw-plugin

fancyboi999 commented Apr 9, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

fancyboi999 commented Apr 9, 2026

Uh oh!

fancyboi999 commented Apr 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		from agentscope.message import Msg, TextBlock
		from agentscope.tool import ToolResponse

Conversation

fancyboi999 commented Apr 9, 2026

背景

做了什么

当前集成范围

本地验证

配套 PR

想讨论的方向

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

fancyboi999 commented Apr 9, 2026

Uh oh!

fancyboi999 commented Apr 10, 2026

两个框架的记忆体系差得挺远

MemOS 能帮 CoPaw 做什么

另一个想法：支持本地模式

几个拿不准的

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants