Bug type

Bug

Summary

Active Memory plugin is properly configured, enabled, and shows as `loaded` / `activated` with 1 hook registered, but does not inject any `<active_memory_plugin>` context into conversational sessions. The agent sees no memory injection at all despite the plugin appearing fully operational.

Environment

• OpenClaw version: 2026.4.10 (44e5b62)
• OS: Linux 5.10.134 (x64), container environment
• Channel: Discord direct message
• Model: glm-turbo (self/glm-turbo)

Steps to Reproduce

1. Configure active-memory per official docs:

{
  "plugins": {
    "entries": {
      "active-memory": {
        "enabled": true,
        "config": {
          "enabled": true,
          "agents": ["main"],
          "allowedChatTypes": ["direct"],
          "modelFallbackPolicy": "default-remote",
          "queryMode": "recent",
          "promptStyle": "balanced",
          "timeoutMs": 15000,
          "maxSummaryChars": 220,
          "persistTranscripts": false,
          "logging": true
        }
      }
    }
  }
}

2. Restart gateway with `openclaw gateway --force`
3. Verify plugin is loaded:
   • `openclaw plugins list` → shows active-memory as "loaded"
   • `openclaw plugins info active-memory` → status: loaded, hookCount: 1, activated: true
4. Send a message in a Discord DM session
5. Check the agent reply — no `<active_memory_plugin>` tags injected into system context

Expected Behavior

The Active Memory plugin should:

1. Run its blocking memory sub-agent before the main reply
2. Inject relevant memory as hidden `<active_memory_plugin>...` context
3. Show verbose output like `🧩 Active Memory: ok 842ms recent 34 chars` when `/verbose on` is enabled

Actual Behavior

• Plugin shows as loaded and activated
• hookCount: 1 (before_prompt_build hook)
• No memory injection occurs in conversational replies
• Agent system context contains no active memory tags
• No verbose output lines for Active Memory

Observations

• `openclaw plugins info active-memory --json` shows `hookNames: []` despite `hookCount: 1` — this may indicate the hook is registered internally but not properly exposed
• The model used is `glm-turbo` (a non-mainstream model) — the Active Memory sub-agent may inherit this model and fail silently (returning NONE)
• `config.validate` passes without errors
• Memory search itself works fine (manual `memory_search` tool calls return results)

Possible Root Causes

1. Model incompatibility: Active Memory sub-agent inherits `glm-turbo` which may not follow the expected response format, causing it to always return `NONE`
2. Hook not firing: `hookNames` is empty despite `hookCount: 1` — the `before_prompt_build` hook may not be properly wired
3. Chat type mismatch: The Discord direct session may not be classified as `"direct"` in the plugin's runtime check
4. Session type filtering: The plugin doc says it only runs on "eligible interactive persistent chat sessions" — there may be additional runtime conditions not documented