decouple scenario, add structure to memories

This commit is contained in:
2026-04-11 22:00:25 +05:30
parent cf6653afbd
commit 48f02e7d44
7 changed files with 689 additions and 122 deletions

435
engine.py Normal file
View File

@@ -0,0 +1,435 @@
from datetime import datetime, timedelta
import logging
import multiprocessing
from langchain_community.chat_models import ChatLlamaCpp
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_core.messages import HumanMessage, SystemMessage
DEFAULT_MODEL_PATH = "/home/sortedcord/.cache/huggingface/hub/models--ggml-org--gemma-4-E4B-it-GGUF/snapshots/6b352c53e1d2e4bb974d9f8cafcf85887c224219/gemma-4-e4b-it-Q4_K_M.gguf"
logger = logging.getLogger(__name__)
embeddings = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")
llm = ChatLlamaCpp(
temperature=0.2,
model_path=DEFAULT_MODEL_PATH,
n_ctx=4096,
n_gpu_layers=8,
max_tokens=512,
n_threads=multiprocessing.cpu_count() - 1,
repeat_penalty=1.5,
)
def _format_prompt(messages):
formatted = []
for message in messages:
formatted.append(f"{message.__class__.__name__}:\n{message.content}")
return "\n\n".join(formatted)
def _normalize_llm_output(text: str) -> str:
return text.replace("\r", "").replace("\n", "").strip()
def _time_of_day_label(hour: int, *, for_today: bool) -> str:
if 5 <= hour < 12:
return "morning"
if 12 <= hour < 17:
return "afternoon"
return "tonight" if for_today else "night"
def describe_relative_time(
timestamp_str: str,
reference_time: datetime,
*,
prefer_day_part_for_today: bool = False,
) -> str:
try:
timestamp = datetime.strptime(timestamp_str, "%Y-%m-%d %H:%M")
except ValueError:
return "a long time ago"
delta = reference_time - timestamp
seconds = delta.total_seconds()
if seconds < 0:
return "just now"
if not prefer_day_part_for_today:
if seconds < 120:
return "just now"
if seconds < 15 * 60:
return "a few minutes ago"
if seconds < 90 * 60:
return "an hour ago"
if seconds < 3 * 60 * 60:
return "a couple hours ago"
day_diff = (reference_time.date() - timestamp.date()).days
if day_diff == 0:
return f"today {_time_of_day_label(timestamp.hour, for_today=True)}"
if day_diff == 1:
return f"yesterday {_time_of_day_label(timestamp.hour, for_today=False)}"
if day_diff == 2:
return "2 days ago"
if day_diff == 3:
return "3 days ago"
if day_diff <= 6:
return "a couple days ago"
if day_diff <= 10:
return "a week ago"
if day_diff <= 20:
return "a couple weeks ago"
if day_diff <= 45:
return "a month ago"
if day_diff <= 75:
return "a couple months ago"
if day_diff <= 420:
return "a year ago"
return "a long time ago"
class WorldClock:
def __init__(self, start_year=1999, month=5, day=14, hour=18, minute=0):
# We use a standard datetime object for easy math
self.current_time = datetime(start_year, month, day, hour, minute)
def advance_time(self, minutes=0, hours=0, days=0):
self.current_time += timedelta(minutes=minutes, hours=hours, days=days)
def get_time_str(self):
# 1999-05-14 18:00
return self.current_time.strftime("%Y-%m-%d %H:%M")
def get_vibe(self):
"""Helper to tell the LLM the 'feel' of the time."""
hour = self.current_time.hour
if 5 <= hour < 12:
return "Morning"
if 12 <= hour < 17:
return "Afternoon"
if 17 <= hour < 21:
return "Evening"
return "Night"
@classmethod
def from_time_str(cls, time_str: str | None):
if not time_str:
return cls()
parsed = datetime.strptime(time_str, "%Y-%m-%d %H:%M")
return cls(
start_year=parsed.year,
month=parsed.month,
day=parsed.day,
hour=parsed.hour,
minute=parsed.minute,
)
class MemoryEntry:
def __init__(self, content, event_type, timestamp_str, location, entities):
self.content = content
self.event_type = event_type # 'dialogue', 'observation', 'reflection'
self.timestamp = timestamp_str
self.location = location
self.entities = entities
def __repr__(self):
return f"[{self.timestamp}] ({self.location}): {self.content}"
def to_dict(self):
return {
"content": self.content,
"event_type": self.event_type,
"timestamp": self.timestamp,
"location": self.location,
"entities": list(self.entities),
}
def to_vector_text(self):
entities = ", ".join(self.entities) if self.entities else "Unknown"
return (
f"{self.content}\n"
f"Time: {self.timestamp}\n"
f"Location: {self.location}\n"
f"Entities: {entities}\n"
f"Type: {self.event_type}"
)
def to_relative_string(self, reference_time: datetime):
time_label = describe_relative_time(self.timestamp, reference_time)
return f"[{time_label}] ({self.location}): {self.content}"
class EntityMemory:
def __init__(self):
self.vector_store = None
self.entries = []
def save(self, entry: MemoryEntry):
self.entries.append(entry)
entry_text = entry.to_vector_text()
if self.vector_store is None:
self.vector_store = FAISS.from_texts(
[entry_text],
embeddings,
metadatas=[{"entry_index": len(self.entries) - 1}],
)
else:
self.vector_store.add_texts(
[entry_text],
metadatas=[{"entry_index": len(self.entries) - 1}],
)
def retrieve(self, query: str, k=2, reference_time: datetime | None = None):
if self.vector_store is None:
return "No long-term memories relevant."
docs = self.vector_store.similarity_search(query, k=k)
memories = []
for doc in docs:
entry_index = doc.metadata.get("entry_index")
if entry_index is None:
memories.append(doc.page_content)
continue
entry = self.entries[entry_index]
if reference_time is None:
memories.append(repr(entry))
else:
memories.append(entry.to_relative_string(reference_time))
return "\n".join(memories)
def dump_entries(self):
return list(self.entries)
class Entity:
def __init__(
self,
name,
traits,
stats,
voice_sample,
current_mood="Neutral",
entity_id=None,
):
self.name = name
self.traits = traits
self.stats = stats
self.current_mood = current_mood
self.memory = EntityMemory()
# TIER 1: The Short-Term Buffer (Verbatim)
self.chat_buffer = []
self.voice_sample = voice_sample
self.entity_id = entity_id
def perceive(self, entry: MemoryEntry):
self.memory.save(entry)
def reflect_and_summarize(self, world_clock: WorldClock, location: str):
"""Converts Tier 1 (Buffer) into Tier 2 (Long-term Subjective Memory)."""
if not self.chat_buffer:
return
dialogue_text = "\n".join(
[f"{m['role_name']}: {m['content']}" for m in self.chat_buffer]
)
# The Subjective Filter Prompt
summary_prompt = [
SystemMessage(
content=f"""
You are the private inner thoughts of {self.name}.
Traits: {", ".join(self.traits)}.
Mood: {self.current_mood}.
Voice Reference: {self.voice_sample}
Think about what just happened.
- No META-TALK, Do not use 'player', 'interaction', 'entity', or 'dialogue'
- BE SUBJECTIVE. If you hated the talk or loved it, then express that.
- USE YOUR VOICE. Match the style of your Voice Reference
- Focus only on facts learned or feelings toward the person"""
),
HumanMessage(
content=f"""
What just happened? Context:\n{dialogue_text}"""
),
]
logger.info("LLM prompt (reflection):\n%s", _format_prompt(summary_prompt))
summary = _normalize_llm_output(llm.invoke(summary_prompt).content)
logger.info("SYSTEM: %s reflected on the talk: '%s'", self.name, summary)
chat_entities = sorted(
{
m["role_id"]
for m in self.chat_buffer
if m.get("role_id") and m.get("role_id") != self.entity_id
}
)
reflection = MemoryEntry(
content=f"Past Conversation Summary: {summary}",
event_type="reflection",
timestamp_str=world_clock.get_time_str(),
location=location,
entities=chat_entities,
)
self.perceive(reflection)
self.chat_buffer = [] # Clear buffer after archiving
class Player(Entity):
pass
def ask_entity(
entity: Entity,
player: Entity,
player_query: str,
world_clock: WorldClock,
location: str,
):
facts = entity.memory.retrieve(
player_query,
reference_time=world_clock.current_time,
)
recent_context = "\n".join(
[f"{m['role_name']}: {m['content']}" for m in entity.chat_buffer[-5:]]
)
world_time_label = describe_relative_time(
world_clock.get_time_str(),
world_clock.current_time,
prefer_day_part_for_today=True,
)
prompt = [
SystemMessage(content=f"WORLD TIME: {world_time_label}"),
SystemMessage(
content=f"""
### ROLE
You are {entity.name}. Persona: {", ".join(entity.traits)}.
Current Mood: {entity.current_mood}.
Vibe Time: {world_clock.get_vibe()}.
Location: {location}.
### WRITING STYLE RULES
1. NO META-TALK. Never mention "memory," "records," "claims," or "narratives."
2. ACT, DON'T EXPLAIN. If you don't know something, just say "Never heard of it" or "I wasn't there." Do not explain WHY you don't know.
### KNOWLEDGE
MEMORIES: {facts}
RECENT CHAT: {recent_context}
"""
),
HumanMessage(content=f"{player.name} speaks to you: {player_query}"),
]
logger.info("LLM prompt (dialogue):\n%s", _format_prompt(prompt))
response = _normalize_llm_output(llm.invoke(prompt).content)
entity.chat_buffer.append(
{
"role_id": player.entity_id,
"role_name": player.name,
"content": player_query,
}
)
entity.chat_buffer.append(
{
"role_id": entity.entity_id,
"role_name": entity.name,
"content": response,
}
)
player.chat_buffer.append(
{
"role_id": player.entity_id,
"role_name": player.name,
"content": player_query,
}
)
player.chat_buffer.append(
{
"role_id": entity.entity_id,
"role_name": entity.name,
"content": response,
}
)
logger.info("[%s]: %s", entity.name.upper(), response)
def _build_name_lookup(entities):
name_lookup = {}
for entity_key, entity in entities.items():
name_lookup[entity_key.lower()] = entity_key
name_lookup[entity.name.lower()] = entity_key
return name_lookup
def start_game(entities, player_id=None, world_time=None, location="Unknown"):
player = None
if player_id:
player = entities.get(player_id)
if player is None:
raise ValueError(f"Player entity '{player_id}' not found in scenario.")
else:
player = Player(
name="Player",
traits=["Curious"],
stats={},
voice_sample="Voice: 'Direct and concise.'",
entity_id="player",
)
available_entities = {
entity_id: entity
for entity_id, entity in entities.items()
if entity_id != player_id
}
world_clock = WorldClock.from_time_str(world_time)
current_entity = None
name_lookup = _build_name_lookup(available_entities)
entity_names = "/".join(
[entity.name for entity in available_entities.values()] + ["Exit"]
)
logger.info("--- WORLD INITIALIZED ---")
logger.info("World initialized with %s active entities.", len(available_entities))
logger.info("Current location: %s", location)
logger.info(
"World time: %s (%s)", world_clock.get_time_str(), world_clock.get_vibe()
)
while True:
target_name = (
input(f"\nWho do you want to talk to? ({entity_names}): ").lower().strip()
)
if target_name in ["exit", "quit"]:
if current_entity:
current_entity.reflect_and_summarize(world_clock, location)
break
target_key = name_lookup.get(target_name)
if target_key is None:
logger.warning("Target not found.")
continue
new_entity = available_entities[target_key]
if current_entity and current_entity != new_entity:
logger.info(
"You leave %s and approach %s.", current_entity.name, new_entity.name
)
current_entity.reflect_and_summarize(world_clock, location)
current_entity = new_entity
user_msg = input(f"You to {current_entity.name}: ")
ask_entity(current_entity, player, user_msg, world_clock, location)