feat(personalization): Personalization pipeline with Mem0 memory integration and terminal chat mode

[RyamL1221]

feat: Personalization pipeline with Mem0 memory integration and terminal chat mode

Summary

This PR adds a personalization pipeline to AIOS that enables context-aware LLM responses based on prior interactions. It includes memory storage/retrieval via Mem0, ChromaDB persistence infrastructure, dynamic Ollama model discovery, and a new terminal chat mode with intelligent intent routing.

Changes

1. Personalization Pipeline (aios/memory/, aios/syscall/)

• ContextInjector (aios/memory/context_injector.py): Retrieves relevant memories from Mem0 before each chat LLM call and prepends them as a system message. Filters by relevance score and respects a configurable token budget.
• ConversationExtractor (aios/memory/conversation_extractor.py): After each chat LLM call, stores the user+assistant exchange as a memory via a background daemon thread. Never blocks the LLM response.
• Integration (aios/syscall/syscall.py): Wired into SyscallExecutor.execute_request() for action_type="chat" queries. ContextInjector runs before the LLM call, ConversationExtractor runs after. Personalization is only active in chat mode — file operations (action_type="operate_file") bypass the pipeline entirely.
• Kernel wiring (runtime/launch.py): Personalization components initialized during initialize_components() when memory.provider is "mem0".

2. Mem0 Provider Enhancements (aios/memory/providers/mem0.py)

• ChromaDB persistence: Injects a chromadb.PersistentClient to bypass Mem0 v1.0.11's deprecated chromadb.Client(Settings(...)) which is always in-memory in ChromaDB 1.5.5. Resolves paths to absolute, creates the persistence directory automatically.
• Default path: When no path is configured for ChromaDB, defaults to .mem0/chroma relative to the project root.
• Known limitation: Cross-restart memory recall doesn't work with small local Ollama models (e.g., qwen2.5:7b) because Mem0's fact extraction produces empty results. In-session personalization works correctly. The data suggests that larger models with stronger instruction-following (e.g., GPT-4, Claude) may produce better fact extraction results, but this has not been tested.

3. Terminal Chat Mode (aios/terminal/, scripts/run_terminal.py, runtime/run_terminal.py)

• IntentRouter (aios/terminal/intent_router.py): Two-stage intent classifier — keyword heuristic (zero latency) with LLM fallback for ambiguous inputs. Classifies user input as chat or file operation.
• Three modes: /auto (default, intelligent routing), /chat (all input → llm_chat() with personalization), /file (all input → llm_operate_file()).
• Conversation history: Chat mode maintains session history so the LLM has context from prior messages.
• Mode indicator: Prompt shows [auto], [chat], or [file] to indicate current mode.
• Both scripts/run_terminal.py and runtime/run_terminal.py updated with identical routing behavior.

4. Kernel Stability (runtime/launch.py)

• Fixed double-initialization bug: changed uvicorn.run("runtime.launch:app") to uvicorn.run(app) and added _ensure_initialized() guard. Components now initialize exactly once.

5. Configuration & CI

• aios/config/config.yaml.example: Added qwen3:4b model, path option for ChromaDB, personalization settings documentation.
• requirements.txt: Added hypothesis for property-based testing.
• .github/workflows/test-ollama.yml: CI pulls both qwen3:1.7b and qwen3:4b.
• .github/workflows/test-qdrant.yml: Minor CI updates.
• .gitignore: Added .mem0/ directory.

New Files

File	Purpose
aios/memory/context_injector.py	Memory retrieval and injection before LLM calls
aios/memory/conversation_extractor.py	Async conversation storage after LLM calls
aios/terminal/__init__.py	Terminal package exports
aios/terminal/intent_router.py	Keyword + LLM intent classification
tests/modules/llm/ollama/test_dynamic_registration_pbt.py	7 property-based tests for dynamic Ollama registration
tests/modules/memory/__init__.py	Test package init

Testing

• 7 property-based tests (Hypothesis) for Ollama dynamic registration — all pass
• 4/4 concurrent Ollama CI tests pass (previously 2 failed)
• 4/4 single Ollama CI tests pass
• In-session personalization verified: name recall, preference tracking, conversation context
• Terminal chat/file/auto mode switching verified with live kernel

Known Limitations

• Cross-restart memory persistence requires models with strong instruction-following for Mem0's fact extraction step. Small local models (qwen2.5:7b) produce empty extraction results. The persistence infrastructure is in place and the data suggests larger models may resolve this, but cross-restart recall has not been verified with cloud-hosted models.
• Conversation history in chat mode grows unbounded within a session. Future improvement: sliding window or token budget.
• Redis connection errors in logs (Error 61 connecting to localhost:6379) are pre-existing — the LSFS file watcher tries to connect to Redis which isn't running locally.