Add support for Hermes model family (e.g., Hermes-3, Nous-Hermes) as a secondary reasoning engine with cross-session memory and adjustment capabilities.

## Summary

Add support for Hermes model family (e.g., Hermes-3, Nous-Hermes) as a secondary reasoning engine with cross-session memory and adjustment capabilities.

## Component(s) Affected

- [x] Core / Backend (app, agents, config, providers, utils, local_models)
- [x] Console (frontend web UI)
- [ ] Channels (DingTalk, Feishu, QQ, Discord, iMessage, etc.)
- [ ] Skills
- [ ] CLI
- [x] Documentation (website)
- [ ] Tests
- [ ] CI/CD
- [ ] Scripts / Deploy

## Problem / Motivation

Users may want to leverage multiple reasoning models for different tasks. Currently:

- **Limited model diversity**: QwenPaw primarily supports Qwen model variants
- **No cross-session learning**: Models don't remember or adapt from past interactions
- **Missing Hermes capabilities**: Hermes models (Nous-Hermes series) are known for:
- Strong reasoning and instruction following
- Excellent performance on complex tasks
- Different strengths compared to Qwen models

Users benefit from:
- Model variety for different use cases
- Cross-session memory (model remembers and adjusts after each session)
- Ability to switch models per chat or task

## Proposed Solution

1. **Add Hermes Model Support**:
- Integrate Hermes-3 and Nous-Hermes model families
- Support via Ollama, LM Studio, and local llama.cpp backends
- Add to Console model selection UI

2. **Cross-Session Memory System**:
- Persistent memory that survives session restarts
- Model can remember user preferences, past decisions, and conversation patterns
- Adjust behavior based on historical interactions
- Integrate with existing ReMe (long-term memory) system

3. **Model Switching**:
- Per-chat model selection
- Easy switching between Qwen and Hermes models
- Preserve context when switching

4. **Documentation**:
- Add Hermes models to model configuration docs
- Provide setup guides for local deployment
- Document memory/adjustment capabilities

## Alternatives Considered

- **Qwen-only**: Current approach; limits model diversity
- **Manual context injection**: Users paste previous conversation summaries, but this is cumbersome
- **External memory tools**: Use separate systems for memory, but loses integration with agent loop

## Additional Context

- Hermes models: https://huggingface.co/NousResearch
- QwenPaw Local runtime already supports llama.cpp; Hermes integration should be feasible
- Cross-session memory aligns with QwenPaw's ReMe v0.4.0 long-term memory system
- Similar to how some platforms support multiple model providers (OpenAI, Anthropic, etc.)

**Benefits**:
- Users get more model options for different tasks
- Cross-session learning improves agent performance over time
- Competitive with other AI platforms that support multiple models

## Willing to Contribute

- [ ] I am willing to open a PR for this feature (after discussion).

Start by tracing the existing ReMe v0.4.0 long-term memory system, the local_models and provider components, and the Console model-selection flow. Compare current Qwen support with the proposed Ollama, LM Studio, and llama.cpp paths; done means Hermes models can be selected per chat, memory persists across sessions, and the configuration and setup documentation is updated.

Vom Indexierungsmodell aus dem Issue-Text verfasst.