Fixes.

server/memory_manager.py

@@ -19,7 +19,6 @@ from .config import settings
 from .logger import logger
 
 # This block is for static analysis only (for Pylance).
-# It is not executed at runtime, so it won't cause an ImportError.
 if TYPE_CHECKING:
     from chonkie import SemanticChunker
     from chonkie.embeddings import SentenceTransformerEmbeddings
@@ -70,7 +69,11 @@ class MemoryManager:
                     raise ImportError("Chonkie is configured but not installed correctly. Please run 'uv pip install \"chonkie[st]\"'.")
                 
                 logger.info("Initializing Chonkie with SemanticChunker...")
-                embedding_handler = SentenceTransformerEmbeddings(model_name=settings.memory.embedding_model, device="cpu")
+                # FIXED: Changed keyword from 'model_name' to 'model'
+                embedding_handler = SentenceTransformerEmbeddings(
+                    model=settings.memory.embedding_model, 
+                    device="cpu"
+                )
                 self.chunker = SemanticChunker(embedding_model=embedding_handler)
                 self.embedder = embedding_handler
                 logger.info("Using 'chonkie' for text chunking.")
@@ -107,9 +110,8 @@ class MemoryManager:
         if not text or not text.strip(): return
         try:
             if settings.memory.chunker == "chonkie" and self.chunker:
-                # The runtime code knows what self.chunker(text) returns
                 chonkie_chunks = self.chunker(text)
-                chunks = [c.text for c in chonkie_chunks if c.text.strip()] # type: ignore (This might break, we're see)
+                chunks = [c.text for c in chonkie_chunks if c.text.strip()] # type: ignore (This *might* break things)
             else:
                 chunks = self._simple_chunker(text)
 

settings.yml

@@ -20,7 +20,7 @@ llm:
   # The model identifier for the chosen backend.
   # Example for OpenRouter: "google/gemma-2-9b-it"
   # Example for LM Studio: "gemma-2-9b-it-gguf" (or whatever you have loaded)
-  story_model: "google/gemma-3n-e4b"
+  story_model: "gemma-3-12b-it-GGUF"
 
   # The format for sending player actions to the AI.
   # "json": (Recommended) More reliable and less ambiguous for modern models.
@@ -62,7 +62,7 @@ audio:
 memory:
   # The chunking strategy to use for breaking down text for the RAG system.
   # Options: "simple" (original method), "chonkie" (advanced library).
-  chunker: "simple"
+  chunker: "chonkie"
 
   # The SentenceTransformer model used to create embeddings for the AI's
   embedding_model: "google/embeddinggemma-300m"