Replies: 2 comments
-
|
把代码看了两遍,横着跟 Codex、Claude Code、Hermes、Mem0、Generative Agents 对了一遍,有四个具体意见。 1. Motive Formation 放在 Recall 之前,这个顺序值得 challengeissue 自己就问了这一点。当前 A-MEM(Xu 2025)的做法是先用结构信号(关键词、tags、similarity links)做预处理,再让 LLM 做高阶判断。我们 Step 0 的 Context Materializer 已经在做类似的事( 建议把 Context Materializer 移到 motive 前面强制跑,把 motive LLM 的输入从"100 条线性列表"换成"N 个 Context 簇 + 未归簇尾巴"。LLM 在已有结构上做判断比从零提取模式简单得多。 2. 缺一个独立的 Prune 阶段Codex Phase 2 之后有显式 我们 pipeline 五阶段里 ARCHIVE 只是 Persistence 里的一个分支,完全依赖 reasoning 阶段的 LLM 主动想到。如果用户最近没新增对应主题的记忆,某条 6 个月前过期的 InsightMemory 永远不会被想到归档,这是系统性盲区。
3. 类型系统不符合 CoALA 公约CoALA(Sumers et al. 2023)给 language agent 立了四种记忆类型公约:working / episodic / semantic / procedural。这是过去两年学术界和 LangMem 等主流框架都在遵循的命名。 我们自创了 5 种:InsightMemory / LongTermMemory / SkillMemory / ProfileMemory / PreferenceMemory。映射一下:
五种类型的边界在代码里其实是模糊的(types.py 里很多地方互相可以替代)。建议中期把类型系统对齐到 CoALA 公约,InsightMemory 保留作为 Dream 专属类型即可。这件事不是为了"看起来正统",而是后面跟其他框架(LangGraph、Letta、Mem0)互通时会少很多翻译成本。 4. 评测短板:我们目前没有任何公开 benchmark 数据LoCoMo(ACL 2024,Maharana et al.)32 sessions / 600 turns / 16K tokens,LongMemEval(2026)各家对比,DMR(MemGPT 团队建立)Zep 94.8% > MemGPT 93.4%。Mem0 在 LOCOMO 上 latency 比 full-context 降 91%。 我们 Dream CE 目前对外的所有讨论都是定性的:"动机驱动"、"假设演绎"、"跨对话模式"。如果不跑一遍上面任一基准,我们连"Dream 比不做 Dream 好"这一点都没有数据支撑。 建议立一个 follow-up issue 总结:三个具体动作 + 一个评测建议。把 Context Materializer 前置、加 Prune 阶段、类型对齐 CoALA、跑 LoCoMo。 |
Beta Was this translation helpful? Give feedback.
-
|
This issue has been automatically marked as stale due to inactivity. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Dream CE is an early community edition module for offline memory consolidation in MemOS. We would like to invite architecture feedback before turning more ideas into implementation tasks.
Current architecture
Dream CE currently follows a staged pipeline:
Main components:
MotiveFormation: groups pending memories into dream motives.DirectRecall: recalls related memories around each motive.ConsolidationReasoning: produces Dream actions and insights.StructuredDiarySummary: generates a human-readable diary artifact.DreamPersistence: persists Dream-produced memories and diary entries.DreamSignalStore: accumulates trigger signals.Design intent
The current implementation is intentionally modular. Community contributors should be able to replace or extend:
Questions
Desired outcome
We want feedback on the architecture before expanding Dream CE further. Concrete proposals may be converted into follow-up
help wantedissues.Beta Was this translation helpful? Give feedback.
All reactions