Project 4: Building an Agent Memory Store

Your challenge is to take a simple CLI chatbot and give it a "Memory Bank." The agent should remember a user's name, their favorite programming language, and the project they were working on three days ago.

---

1. Project Requirements

• Agent Framework: Simple Python input() loop or a LangGraph agent.
• Memory Type: Mix of Episodic (Recent chat logs) and Semantic (Extracted facts).
• Consolidation: A daily "Summarizer" that turns chat logs into persistent facts.

---

2. Logic Flow

1. User Input: "I'm move from Python to Go for my latest project."
2. Action: The agent stores this interaction (Episodic).
3. Trigger: Every 5 turns, the agent calls an LLM to "Extract Facts."
4. New Memory: "Preference: Learning Go," "Current Task: Project Migration."
5. Retrieval: Next time the user asks "What should I do next?", the agent retrieves these facts and suggests Go-related tasks.

---

3. Implementation (Python)

def consolidate_memory(session_id, collection):
    # 1. Get recent chat logs
    recent_logs = collection.query(
        where={"session_id": session_id, "type": "episodic"},
        n_results=10
    )
    
    # 2. Extract Facts via LLM
    text_to_analyze = " ".join(recent_logs['documents'][0])
    facts = llm.extract_facts(text_to_analyze)
    
    # 3. Store as Semantic Truths
    for fact in facts:
        collection.add(
            documents=[fact],
            metadatas=[{"type": "semantic", "confidence": 0.9}],
            ids=[generate_uid()]
        )

---

4. Evaluation Criteria

• Knowledge Retrieval: After telling the agent a fact, can it recall it 20 turns later (after the context window has cleared)?
• Fact Accuracy: Does the consolidation process correctly extract the "Truth" or does it misinterpret the conversation?
• Speed: Does the "Recall" step significantly delay the chat response?

---

Deliverables

1. A Python script where the agent demonstrates memory across two different execution runs.
2. A technical diagram of your "Memory Consolidation" pipeline.

---

Building agents with memory is the key to creating AI that feels truly intelligent. Good luck!