ctxovrflw Memory Plugin for OpenClaw

Replace OpenClaw's built-in memory with ctxovrflw — a local-first, privacy-focused AI memory layer with semantic search, knowledge graph, and cross-tool recall.

What It Does

• Replaces memory_search / memory_get with ctxovrflw's semantic search (hybrid keyword + vector)
• Adds memory_store to persist important context across sessions
• Auto-recall: Automatically injects relevant memories into agent context before each turn
• Auto-capture: Optionally stores important user messages as memories
• Knowledge graph: Graph-boosted recall surfaces related entities and connections
• Cross-tool: Memories stored by Cursor, Claude Code, Cline, or any MCP-connected tool are available in OpenClaw
• /ctxovrflw command: Quick status check without invoking the LLM

Prerequisites

1. Install ctxovrflw: curl -fsSL https://ctxovrflw.dev/install.sh | bash
2. Run init: ctxovrflw init
3. Start daemon: ctxovrflw start

Install

openclaw plugins install @ctxovrflw/memory-ctxovrflw

Tip: ctxovrflw init detects OpenClaw automatically and offers to install the plugin, skill, and agent rules for you via an interactive TUI.

Configure

Add to your OpenClaw config:

{
  plugins: {
    slots: {
      memory: "memory-ctxovrflw"  // Replace built-in memory
    },
    entries: {
      "memory-ctxovrflw": {
        enabled: true,
        config: {
          authToken: "<from ~/.ctxovrflw/config.toml>",
          // Optional:
          daemonUrl: "http://127.0.0.1:7437",  // default
          agentId: "openclaw",                   // default
          autoRecall: true,                      // default
          autoCapture: false,                    // default
          recallLimit: 5,                        // default
          recallMinScore: 0.3,                   // default
          captureMaxChars: 500,                  // default
        }
      }
    }
  }
}

Then restart the gateway:

openclaw gateway restart

How It Works

Auto-Recall (default: on)

Before each agent turn, the plugin searches ctxovrflw for memories relevant to the user's message and injects them as context. The agent sees these automatically — no tool call needed.

Auto-Capture (default: off)

After each agent turn, the plugin scans user messages for important content (preferences, decisions, facts) and stores them automatically. Enable with autoCapture: true.

Tool-Based Access

The agent can also use tools explicitly:

• memory_search — Semantic search with optional subject filter
• memory_store — Store new memories with type, tags, subject
• memory_forget — Delete a memory by ID
• memory_status — Check daemon status, memory count, tier

CLI

openclaw memory status    # Daemon status
openclaw memory search "deployment preferences"
openclaw memory store "Max prefers Railway" --type preference
openclaw memory subjects  # List all subjects

vs Built-in Memory

| Feature | memory-core | memory-lancedb | memory-ctxovrflw | |---|---|---|---| | Storage | Markdown files | LanceDB + OpenAI | SQLite + local ONNX | | Embeddings | None | OpenAI API ($) | Local ONNX (free) | | Privacy | Files on disk | API calls to OpenAI | Everything local | | Cross-tool | No | No | Yes (MCP server) | | Knowledge graph | No | No | Yes (Pro) | | Hybrid search | No | Vector only | Semantic + keyword | | Auto-recall | No | Yes | Yes | | Auto-capture | No | Yes | Yes | | Cloud sync | No | No | Yes (E2E encrypted) |

License

MIT