Bug Description

When context compression triggers (at 100% context), the compacted summary is missing the most recent conversation turns, and fragments from other unrelated sessions are injected into the current context. This causes the agent to lose track of who the user is and what they were just discussing.

Steps to Reproduce

1. Have a multi-turn conversation that approaches context window limit (~100%)
2. Context compression triggers automatically
3. Observe: agent responds as if the conversation just started — "introduce yourself" again, no memory of recent exchanges
4. In some cases, content from completely different sessions appears mixed into the context (e.g., search results from other topics, unrelated conversation fragments)

Actual Behavior

After compression, the agent:

• Loses memory of the most recent N turns (where N varies)
• May have cross-session contamination — content from session A appearing in session B
• Appears to the user as if the conversation just started fresh

The compacted summary appears to be generated from a truncated window that excludes the most recent messages, and/or the merge step incorrectly pulls from wrong session data.

Expected Behavior

After compression:

• All recent conversation context should be preserved in the summary
• No cross-session contamination should occur
• The agent should seamlessly continue the conversation without memory loss

Additional Context

• Hermes version: v0.9.0 (2026.4.13)
• Model: MiniMax-M2.7-highspeed
• Platform: Feishu (websocket mode)
• This has happened at least 2 times in the same day for the reporter

Related Issues

• Context compression causes historical conversations to be inserted into current conversation #9096 mentions historical conversations being inserted, but the core issue here is more severe: recent conversation is lost AND cross-session contamination occurs simultaneously
• [Bug]: Context compression causes incoherent responses on small-context models #7133 mentions incoherent responses on small-context models