Adds Long-Term-memory to microchain agent (#129)

* Adding sqlalchemy support

* Added long term storage

* Using SQLTable for db interactions

* Added test for db_storage.py | tested microchain manually

* Updated poetry.lock

* Minor improvements

* Updated poetry.lock

* Fixing mypy etc

* Added PR comments

* Changes after merge

* Updated poetry.lock

* Fixing mypy

* Fixed isort

* Implemented PR comments

* Fixed isort

* Fixed black

prediction_market_agent/agents/microchain_agent/app.py

@@ -11,7 +11,7 @@
 from prediction_market_agent_tooling.tools.costs import openai_costs
 from prediction_market_agent_tooling.tools.utils import check_not_none
 
-from prediction_market_agent.agents.microchain_agent.microchain_agent import get_agent
+from prediction_market_agent.agents.microchain_agent.microchain_agent import build_agent
 from prediction_market_agent.agents.microchain_agent.utils import (
     has_been_run_past_initialization,
 )
@@ -65,7 +65,7 @@ def execute_reasoning(agent: Agent, reasoning: str, model: str) -> None:
 
 # Initialize the agent
 if "agent" not in st.session_state:
-    st.session_state.agent = get_agent(market_type=MarketType.OMEN, model=model)
+    st.session_state.agent = build_agent(market_type=MarketType.OMEN, model=model)
     st.session_state.agent.reset()
     st.session_state.agent.build_initial_messages()
     st.session_state.running_cost = 0.0

prediction_market_agent/agents/microchain_agent/deploy.py

@@ -2,7 +2,7 @@
 from prediction_market_agent_tooling.deploy.agent import DeployableAgent
 from prediction_market_agent_tooling.markets.markets import MarketType
 
-from prediction_market_agent.agents.microchain_agent.microchain_agent import get_agent
+from prediction_market_agent.agents.microchain_agent.microchain_agent import build_agent
 
 
 class DeployableMicrochainAgent(DeployableAgent):
@@ -14,7 +14,7 @@ def run(self, market_type: MarketType) -> None:
         Override main 'run' method, as the all logic from the helper methods
         is handed over to the agent.
         """
-        agent: Agent = get_agent(
+        agent: Agent = build_agent(
             market_type=market_type,
             model=self.model,
         )

prediction_market_agent/agents/microchain_agent/functions.py

@@ -6,6 +6,7 @@
 from prediction_market_agent_tooling.markets.data_models import Currency, TokenAmount
 from prediction_market_agent_tooling.markets.markets import MarketType
 
+from prediction_market_agent.agents.microchain_agent.memory import LongTermMemory
 from prediction_market_agent.agents.microchain_agent.utils import (
     MicroMarket,
     get_balance,
@@ -15,6 +16,7 @@
     get_no_outcome,
     get_yes_outcome,
 )
+from prediction_market_agent.db.models import LongTermMemories
 from prediction_market_agent.tools.mech.utils import (
     MechResponse,
     MechTool,
@@ -297,21 +299,6 @@ def __init__(self, market_type: MarketType) -> None:
         )
 
 
-class SummarizeLearning(Function):
-    @property
-    def description(self) -> str:
-        return "Use this function summarize your learnings and save them so that you can access them later."
-
-    @property
-    def example_args(self) -> list[str]:
-        return [
-            "Today I learned that I need to check my balance fore making decisions about how much to invest."
-        ]
-
-    def __call__(self, summary: str) -> str:
-        return summary
-
-
 class GetBalance(MarketFunction):
     @property
     def description(self) -> str:
@@ -351,10 +338,30 @@ def __call__(self) -> list[str]:
         return [str(position) for position in positions]
 
 
+class RememberPastLearnings(Function):
+    def __init__(self, long_term_memory: LongTermMemory) -> None:
+        self.long_term_memory = long_term_memory
+        super().__init__()
+
+    @property
+    def description(self) -> str:
+        return """Use this function to fetch information about the previous actions you executed. Examples of past 
+        activities include previous bets you placed, previous markets you redeemed from, balances you requested, 
+        market positions you requested, markets you fetched, tokens you bought, tokens you sold, probabilities for 
+        markets you requested, among others.
+        """
+
+    @property
+    def example_args(self) -> list[str]:
+        return []
+
+    def __call__(self) -> t.Sequence[LongTermMemories]:
+        return self.long_term_memory.search()
+
+
 MISC_FUNCTIONS = [
     Sum,
     Product,
-    # SummarizeLearning,
 ]
 
 # Functions that interact with the prediction markets

prediction_market_agent/agents/microchain_agent/memory.py

@@ -0,0 +1,22 @@
+# inspired by crewAI's LongTermMemory (https://github.com/joaomdmoura/crewAI/blob/main/src/crewai/memory/long_term/long_term_memory.py)
+from typing import Any, Dict, Sequence
+
+from prediction_market_agent.db.db_storage import DBStorage
+from prediction_market_agent.db.models import LongTermMemories
+
+
+# In the future, create a base class which this class extends.
+class LongTermMemory:
+    def __init__(self, task_description: str):
+        self.task_description = task_description
+        self.storage = DBStorage()
+
+    def save_history(self, history: list[Dict[str, Any]]) -> None:
+        """Save item to storage. Note that score allows many types for easier handling by agent."""
+        self.storage.save_multiple(
+            task_description=self.task_description,
+            history=history,
+        )
+
+    def search(self, latest_n: int = 5) -> Sequence[LongTermMemories]:
+        return self.storage.load(self.task_description, latest_n)

prediction_market_agent/agents/microchain_agent/microchain_agent.py

@@ -1,4 +1,5 @@
 import typer
+from functions import RememberPastLearnings
 from microchain import LLM, Agent, Engine, OpenAIChatGenerator
 from microchain.functions import Reasoning, Stop
 from prediction_market_agent_tooling.markets.markets import MarketType
@@ -7,6 +8,7 @@
     MARKET_FUNCTIONS,
     MISC_FUNCTIONS,
 )
+from prediction_market_agent.agents.microchain_agent.memory import LongTermMemory
 from prediction_market_agent.agents.microchain_agent.omen_functions import (
     OMEN_FUNCTIONS,
 )
@@ -22,10 +24,11 @@
 """
 
 
-def get_agent(
+def build_agent(
     market_type: MarketType,
     model: str,
     api_base: str = "https://api.openai.com/v1",
+    long_term_memory: LongTermMemory | None = None,
 ) -> Agent:
     engine = Engine()
     engine.register(Reasoning())
@@ -36,6 +39,10 @@ def get_agent(
         engine.register(function(market_type=market_type))
     for function in OMEN_FUNCTIONS:
         engine.register(function())
+
+    if long_term_memory:
+        engine.register(RememberPastLearnings(long_term_memory))
+
     generator = OpenAIChatGenerator(
         model=model,
         api_key=APIKeys().openai_api_key.get_secret_value(),
@@ -58,15 +65,22 @@ def main(
     iterations: int = 10,
     seed_prompt: str | None = None,
 ) -> None:
-    agent = get_agent(
+    # This description below serves to unique identify agent entries on the LTM, and should be
+    # unique across instances (i.e. markets).
+    unique_task_description = f"microchain-agent-demo-{market_type}"
+    long_term_memory = LongTermMemory(unique_task_description)
+
+    agent = build_agent(
         market_type=market_type,
         api_base=api_base,
         model=model,
+        long_term_memory=long_term_memory,
     )
     if seed_prompt:
         agent.bootstrap = [f'Reasoning("{seed_prompt}")']
     agent.run(iterations=iterations)
     # generator.print_usage() # Waiting for microchain release
+    long_term_memory.save_history(agent.history)
 
 
 if __name__ == "__main__":