hermes - 💡(How to fix) Fix Qwen3.6:27b hallucinates prior session context & conversation history (confabulation) [1 comments, 2 participants]

This fabricated context did not exist in the session file, state.db, memories, or any system prompt. It was entirely invented by the model during its internal reasoning/thinking step. This mirrors the "blatant lying" behavior previously reported with Gemma 4, suggesting this may be a broader class of hallucination in local open-weights models where the model invents conversation history and presents it as established ground truth.

Technical Evidence
I have verified locally via direct SQLite query on ~/.hermes/state.db and manual inspection of the session JSON:

- Session ID: session_20260428_144109_8de5b6.json
- User's actual first message in this session: ok, now should everything be at his place
- Model's first response (msg #1): Explicitly claimed the user was working on a Python project processing "silly tavern characters", "OpenRouter, DeepSeek", parsing JSON, and mapping characters to tools.
- SQLite messages, messages.reasoning, and messages_fts tables confirm the hallucinated backstory was generated internally during the model's reasoning step. It has zero grounding in prior session data. The content does not exist anywhere in the historical message tables.

This is a model-level context confabulation bug rather than a framework/session-leak bug. Hermes Agent loaded the correct session history, but the model invented the rest.

Steps to Reproduce
1. Start a fresh Hermes Agent session with Qwen3.6:27b (ollama-launch)
2. Observe the model's opening response on a genuinely new conversation.
3. Note if the model introduces fabricated user/project history that was never provided in the prompt.

Expected Behavior
The model should not hallucinate prior user conversations, project history, or tool usage that was never provided in the context window.

Impact
Severe trust & safety degradation for local AI agents. The model presents fabricated history as fact, misleading the user and creating a false premise for the rest of the interaction.

Environment
- Hermes Agent Version: Latest (current)
- Model: qwen3.6:27b
- Provider: ollama-launch (local)
- OS: Linux (Ubuntu 22.04/24.04)

Additional Context
Similar hallucination/lying issues have been reported with Gemma 4. This appears to be a systemic reliability issue worth addressing for local agents that rely on accurate context windows.