openclaw - 💡(How to fix) Fix [Bug] active-memory recall injects assistant chitchat into <active_memory_plugin> instead of NONE or valid memory summary [2 comments, 2 participants]

[Bug] active-memory recall sometimes injects assistant chitchat into <active_memory_plugin> instead of NONE or a valid memory summary

Summary

The active-memory plugin sometimes injects assistant-style chitchat / clarification boilerplate into runtime context via:

<active_memory_plugin>...</active_memory_plugin>

instead of returning either:

• NONE
• or a valid compact memory summary

This contaminates the main prompt with irrelevant assistant-language such as greetings, “message got cut off” boilerplate, time announcements, and help-offer text.

Expected behavior

When no useful recall is available, the active-memory subagent should return NONE (or no summary after normalization), and nothing should be injected into <active_memory_plugin>.

When useful recall exists, it should return a compact factual memory summary only.

It should never inject:

• greetings
• help-offer boilerplate
• “your message got cut off” clarification text
• time/date announcements
• generic “please tell me what you need help with” text

Actual behavior

The active-memory recall subagent sometimes returns assistant-style free-form text such as:

<active_memory_plugin>
您好！请问有什么可以帮助您的吗？如果您有任何问题或需要帮助，请随时告诉我。
</active_memory_plugin>

Other observed variants include:

• It seems like your message got cut off. Could you please provide more details...
• Hello! It looks like your message didn't come through...
• 当前日期和时间是...如果您需要任何帮助...
• 您好！看起来您可能没有说出您的问题或请求...
• 您好！看起来您的消息可能没有包含具体问题或请求...

These strings are then injected directly into runtime context and can influence downstream prompt behavior.

Why this matters

This is not just low-quality recall. It is prompt contamination:

• assistant-language enters untrusted runtime context
• irrelevant text is injected before the main model turn
• repeated contamination can confuse downstream reasoning
• some variants may propagate into memory artifacts / derived context

Evidence / timeline

Earliest known subagent drift

• 2026-04-22T09:12:12.009Z
• file:
  • /root/.openclaw/plugins/active-memory/transcripts/agents/main/active-memory/active-memory-mo9u3b5u-3b670fb0.jsonl
• model:
  • bigmodel/glm-4-flash-250414
• observed assistant output:
  • 你好！请问有什么可以帮助您的？如果您有任何问题或需要帮助，请随时告诉我。

Earliest known main-session contamination

• starting at 2026-04-28T00:26:05.715Z
• file:
  • /root/.openclaw/agents/main/sessions/9d4f2cf8-d54d-4bba-a695-d1c9ff5dd9f3.jsonl.bak-2045903-1777341207869

Observed injected variants included:

• Hello! It looks like you didn't finish your message...
• Hello! It seems like your message got cut off...
• 你好！请问有什么可以帮助你的吗？...
• 您好！请问有什么可以帮助您的吗？...

These match the style of raw active-memory subagent outputs.

Root cause hypothesis

The most likely root cause is:

1. the active-memory recall subagent is instructed to return either:
   • NONE
   • or one compact plain-text summary
2. in weak-match / no-recall cases, some models (especially observed with bigmodel/glm-4-flash-250414) do not reliably follow that contract
3. instead of returning NONE, they produce assistant-style boilerplate
4. the plugin currently treats that output as a valid summary and injects it into <active_memory_plugin>

So the failure is:

• model output contract violation
• plus
• insufficient parser-side validation before injection

Important code-path findings

From local inspection of the installed active-memory bundle:

1) XML packaging layer does not generate the chitchat

buildMetadata(summary) only escapes and wraps the summary:

function buildMetadata(summary) {
  if (!summary) return;
  return [
    `<${ACTIVE_MEMORY_PLUGIN_TAG}>`,
    escapeXml(summary),
    `</${ACTIVE_MEMORY_PLUGIN_TAG}>`
  ].join("\n");
}

So the boilerplate is not being invented there.

2) timeout / partial recovery also does not invent the chitchat

buildTimeoutRecallResult() consumes:

• params.rawReply
• or partial assistant text from transcript

and then passes it through normalizeActiveSummary().

So timeout recovery is also replaying subagent output, not fabricating it.

3) current prompt is still a soft natural-language contract

The recall prompt currently tells the model to return either:

• NONE
• or one compact plain-text summary

That is a soft instruction, not a strict protocol.

Configuration correlation

A likely behavioral regression coincides with an active-memory recall model switch.

Observed historical configs:

before switch

openclaw.json.bak_before_am_switch

{
  "model": "doubao/doubao-1-5-lite-32k-250115",
  "timeoutMs": 20000,
  "maxSummaryChars": 600,
  "persistTranscripts": true,
  "queryMode": "recent",
  "promptStyle": "balanced"
}

after switch

openclaw.json.bak_am_optimize

{
  "model": "bigmodel/glm-4-flash-250414",
  "timeoutMs": 20000,
  "maxSummaryChars": 600,
  "persistTranscripts": true,
  "queryMode": "recent",
  "promptStyle": "balanced"
}

This suggests the model switch may have increased invalid non-NONE outputs in weak/no-recall cases.

Recommended fix

Phase 1: minimal hotfix

Harden normalizeActiveSummary() so it fails closed on obvious assistant-chitchat / clarification / time-broadcast output.

This is the smallest safe patch because it protects:

• normal successful recall
• timeout partial recovery
• all current plain-text paths

without requiring a larger redesign.

Phase 2: protocol hardening

Move active-memory recall output to a structured response protocol, e.g.:

{"kind":"none"}

or

{"kind":"memory","summary":"User prefers aisle seats and extra buffer over tight connections."}

Then reject anything that is not valid structured output.

This would make the parser, not the model, the safety boundary.

Minimal patch direction

A minimal practical fix is to validate summary-like output before accepting it, e.g. reject patterns matching:

• greetings + help-offer boilerplate
• “message got cut off”
• “please provide more details”
• generic clarification requests
• time/date broadcast text
• “tell me what help you need” style output

This should be applied inside normalizeActiveSummary() so both normal and timeout-partial paths are covered.

Risk / tradeoffs

False positives

A few legitimate summaries containing similar phrases could be dropped.

This seems acceptable compared with injecting assistant chitchat into every downstream prompt.

False negatives

Pattern-based filtering may still miss future variants.

So Phase 1 is only a stopgap.

Better long-term fix

Structured output + strict parse is more robust than blacklist filtering.

Suggested follow-up

1. add parser-side validation immediately
2. A/B test active-memory recall with:
   • bigmodel/glm-4-flash-250414
   • doubao/doubao-1-5-lite-32k-250115
3. compare:
   • invalid non-NONE rate
   • useful recall rate
   • latency
4. if needed, move to JSON-only recall output contract

Reproduction idea

Try weak / no-recall prompts or ambiguous short inputs under active-memory with queryMode=recent and promptStyle=balanced, then inspect whether the recall subagent returns:

• NONE as expected
• or assistant-style clarification/chitchat that gets injected into <active_memory_plugin>