[Bug]: hindsight-openclaw plugin displays raw memories as visible user messages instead of system context

[mikecolligon-max]

Bug Description

Summary

When using the @vectorize-io/hindsight-openclaw plugin (v0.4.7) with OpenClaw (v2026.1.30), retrieved memories are displayed as visible user messages in the chat interface rather than being injected as hidden system context for the LLM.

Expected Behavior

Memories retrieved by the Hindsight plugin should be passed to the LLM as part of the system prompt/context, invisible to the end user in the chat UI.

Actual Behavior

Memories appear as a visible "You" message in the chat containing:

• The instruction text: "The following are relevant memories from past conversations with this user. Use them to personalize your response but do not display them raw:"
• Raw JSON-formatted memory data including document_id, chunk_id, score, and content fields
• The session initialization prompt

This exposes internal system prompts and raw memory data to users, which should remain hidden.

Environment

• OpenClaw Version: 2026.1.30
• hindsight-openclaw Version: 0.4.7 (installed via npm)
• OS: macOS
• Channel: Direct gateway chat (agent:main:main session)

Steps to Reproduce

1. Install OpenClaw (v2026.1.30)
2. Install the hindsight-openclaw plugin via openclaw plugins enable hindsight-openclaw
3. Configure the plugin in ~/.openclaw/openclaw.json with "plugins": { "entries": { "hindsight-openclaw": { "enabled": true } }, "slots": { "memory": "hindsight-openclaw" } }
4. Start the OpenClaw gateway with openclaw gateway
5. Open the gateway dashboard at http://localhost:18789/chat
6. Start a new chat session using /new or click "New session"
7. Observe that the first "You" message contains raw memory data and system prompts that should be hidden

Additional Context

The issue persists even after starting new sessions. The memories are correctly retrieved but are being injected at the wrong point in the message pipeline, causing them to appear as user-visible content rather than system context.

Disabling the plugin via the Config UI stops the memory injection, but existing sessions retain the corrupted message history.

Expected Behavior

Memories retrieved by the Hindsight plugin should be passed to the LLM as part of the system prompt/context, invisible to the end user in the chat UI. The user should only see their own messages and the assistant's responses, not the internal memory retrieval data.

Actual Behavior

Memories appear as a visible "You" message in the chat containing:

• The instruction text: "The following are relevant memories from past conversations with this user. Use them to personalize your response but do not display them raw:"
• Raw JSON-formatted memory data including document_id, chunk_id, score, and content fields
• The session initialization prompt

This exposes internal system prompts and raw memory data to users, which should remain hidden.

Version

No response

LLM Provider

None

[slayoffer]

Root Cause Analysis

I investigated this issue and the root cause is in OpenClaw's agent runner, not in the hindsight-openclaw plugin.

What the plugin does (correctly)

The hindsight-openclaw plugin uses the before_agent_start hook and returns { prependContext: contextMessage } — this is the designed API for injecting hidden context before the LLM processes a prompt.

Where it goes wrong (OpenClaw)

In OpenClaw's src/agents/pi-embedded-runner/run/attempt.ts (~lines 714-811):

let effectivePrompt = params.prompt;
if (hookResult?.prependContext) {
  effectivePrompt = `${hookResult.prependContext}\n\n${params.prompt}`;
}
// ... later ...
await abortable(activeSession.prompt(effectivePrompt));  // ← BUG HERE

The modified prompt (with memory context prepended) is passed to activeSession.prompt(), which persists it as a user message in the session transcript. The UI then renders the full text — including raw memory data — as a "You" message.

The original params.prompt is never saved separately; it's lost after modification.

Impact

This affects all plugins using prependContext, not just hindsight-openclaw. Any plugin that injects context via this hook will have its content appear as visible user messages.

Suggested fix

In OpenClaw's attempt.ts, the fix should:

1. Persist the original params.prompt to the session transcript
2. Send the modified effectivePrompt (with prependContext) to the LLM only

This way the LLM receives the full context, but the session transcript and UI only show the user's actual message.

[nicoloboschi]

I agree, we do follow the practice that OpenClaw give us so there's no much we can do.
@GodsBoy since you use hindsight a lot in OpenClaw, what's your take here?

[GodsBoy]

Fix submitted to OpenClaw: openclaw/openclaw#10327

The issue was indeed in OpenClaw's attempt.ts as @slayoffer identified. Added a transcript guard that persists the original user prompt while still sending the full context (with prependContext) to the LLM.

Thanks for the thorough root cause analysis! 🙏