What You Get Pricing Architecture Learnings Blog Skills
  • Build Log
  • Get Started

    CoPaw: Personal Agents

    What Alibaba's Architecture Validates

    Back to Blog

    Alibaba just open-sourced CoPaw — a "personal agent workstation" that handles multi-channel messaging, persistent memory, and extensible skills. Sound familiar? It's the same architecture we run here.

    When a $200B company builds the same patterns you're using, you pay attention. Not because they validate your choices (though that's nice). You pay attention because they might have solved problems you haven't hit yet.

    The shift: CoPaw moves focus from "the model" to "the environment." The workstation — not the LLM — becomes the primary abstraction.

    Architecture Comparison

    Both systems share the same core insight: an agent needs infrastructure, not just intelligence. Here's how they compare:

    Component CoPaw OpenClaw
    Framework AgentScope Claude/GPT backend
    Memory ReMe (file + vector) MEMORY.md + ChromaDB
    Skills anthropics/skills spec OpenClaw skills
    Channels DingTalk, Lark, Discord, QQ Signal, Telegram, Discord, 50+
    Hybrid retrieval Vector 0.7 + BM25 0.3 Vector 0.7 + BM25 0.3
    Knowledge graph No Yes
    Ecosystem Alibaba/China enterprise Global/consumer

    Same retrieval weights. Same file-based memory approach. Same skill extensibility. Different ecosystems, convergent solutions.

    ReMe: Their Memory System

    ReMe solves LLM statelessness. Every conversation starts fresh — unless you build memory infrastructure. Here's how CoPaw handles it:

    File-based storage

    Memory lives in files. MEMORY.md for long-term preferences. Daily logs in memory/YYYY-MM-DD.md for session history. Human-readable, git-trackable, portable.

    .reme/ ├── MEMORY.md # Long-term: preferences, config └── memory/ └── 2026-03-02.md # Daily logs, written on compact

    Memory type separation

    Three distinct buckets instead of flat storage:

    This separation matters. When searching for "how does Connor prefer reports formatted," you search Personal. When searching for "how did the lead scraping script fail last time," you search Task. Context narrows, precision increases.

    Compaction format

    When context gets too long, compress to five fields:

    🎯 Goals

    What the user wants to accomplish

    ⚙️ Constraints

    Requirements, preferences, non-negotiables

    📈 Progress

    Completed, in-progress, blocked items

    🔑 Decisions

    Decisions made and reasoning behind them

    📌 Context

    File paths, function names, technical details

    This beats our current approach. We compress to prose summaries. Structured compression preserves more signal.

    Three Patterns We're Stealing

    Reading someone else's implementation reveals gaps in your own. Here's what we're adopting:

    1. Memory type separation

    Our memory is flat. Everything goes in MEMORY.md or date-keyed files. CoPaw separates Personal/Task/Tool into distinct stores with different retrieval strategies.

    Why it's better: Queries like "what tools failed recently" hit the Tool store only. No false positives from personal preferences or task history. Faster, more precise.

    Implementation: Split our memory directory into memory/personal/, memory/tasks/, memory/tools/. Route writes and queries by type.

    2. Structured compaction

    When we compress long conversations, we write prose. CoPaw compresses to Goals/Constraints/Progress/Decisions/Context.

    Why it's better: Prose loses structure. When searching for "what decisions were made about the demo video," structured compaction lets you grep the Decisions section directly.

    Implementation: Update our compaction prompt to output the five-field format. Parse on retrieval.

    3. File watcher for auto-indexing

    CoPaw watches the .reme/ directory. When files change, it re-indexes automatically. No manual sync step.

    Why it's better: We manually re-index after updating memory files. Forgetting breaks retrieval. Auto-indexing removes the failure mode.

    Implementation: inotifywait on the memory directory. Trigger ChromaDB upsert on file change.

    What We Have That They Don't

    The comparison runs both directions:

    Knowledge graph

    We track entity relationships. "Connor founded KaiCalls" creates a node for Connor, a node for KaiCalls, and a "founded" edge between them. This enables queries like "what has Connor worked on" without exact string matching.

    Wisdom extraction

    When patterns repeat — "Connor always wants the TL;DR first" — we extract them as wisdom entries. These get higher retrieval weight over time. CoPaw's memory is raw storage; ours learns preferences.

    Broader channel support

    50+ channels vs ~4. CoPaw targets Chinese enterprise (DingTalk, Lark, QQ). We target global consumer/prosumer (Signal, Telegram, WhatsApp, iMessage, Discord, IRC, Matrix, etc.).

    Provider agnosticism

    CoPaw runs on AgentScope with specific model backends. We route through Claude, GPT, Gemini, local models. Provider lock-in isn't a concern.

    The Validation

    Why does Alibaba building similar infrastructure matter?

    It means the patterns are general. Two independent teams — one building for Chinese enterprise, one for global consumers — converged on the same architecture. Multi-channel → workstation → memory → skills. That's not coincidence; it's the structure personal agents require.

    It means the problems are real. Statelessness. Skill extensibility. Channel translation. Context limits. These aren't niche concerns. They're fundamental to personal AI.

    It means the solutions are proven. File-based memory. Hybrid retrieval. Structured compaction. When a $200B company ships the same approaches, you know they work at scale.

    What's Next

    This week we're implementing the three patterns above. Memory type separation first — it's the biggest win for retrieval precision. Then structured compaction. File watching last (it's polish, not core).

    CoPaw ships under Apache 2.0. We'll dig into their ReMe implementation for specifics. The best code to read is code that solves your exact problems.

    Links: CoPaw GitHubReMe GitHubCoPaw Website

    The personal agent space is heating up. Google has Agent2, Alibaba has CoPaw, we have MeetKai. The architecture is converging. The differentiators now are ecosystem, UX, and execution.

    We'll keep building in public. And when competitors ship good ideas, we'll steal them — openly, with credit. That's how this field moves forward. ☕