┌─────────────────────────────────────────────────────────┐
│                    Agent Runtime                        │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  Session 1          Session 2          Session N        │
│  ┌─────────┐        ┌─────────┐        ┌─────────┐     │
│  │Short-   │        │Short-   │        │Short-   │     │
│  │term     │        │term     │        │term     │     │
│  │Memory   │        │Memory   │        │Memory   │     │
│  └────┬────┘        └────┬────┘        └────┬────┘     │
│       │                  │                  │          │
│       └──────────────────┼──────────────────┘          │
│                          ▼                             │
│              ┌───────────────────────┐                 │
│              │   Reflection Engine   │                 │
│              │  (Background Process) │                 │
│              └───────────┬───────────┘                 │
│                          ▼                             │
│              ┌───────────────────────┐                 │
│              │   Episodic Memory     │                 │
│              │  (Long-term Storage)  │                 │
│              │                       │                 │
│              │  • User prefers X     │                 │
│              │  • Learned fact Y     │                 │
│              │  • Historical event Z │                 │
│              └───────────────────────┘                 │
│                                                         │
└─────────────────────────────────────────────────────────┘

import boto3

control = boto3.client('bedrock-agentcore-control')

**Configure Memory Strategies**:
```python

---

import boto3
import datetime
import uuid

client = boto3.client('bedrock-agentcore')

**Batch Event Storage**:
```python

---

undefined

**Retrieve All Memories for User**:
```python

**Context-Aware Retrieval**:
```python
def get_relevant_memories(memory_id, user_id, current_context):
    """Retrieve memories relevant to current conversation context"""

    # Extract key topics from current context
    topics = extract_topics(current_context)

    # Retrieve for each topic
    all_memories = []
    for topic in topics:
        response = client.retrieve_memory_records(
            memoryId=memory_id,
            actorId=user_id,
            retrievalQuery={
                'semanticQuery': topic,
                'maxRecords': 5
            }
        )
        all_memories.extend(response['memoryRecords'])

    # Deduplicate and rank
    unique_memories = deduplicate(all_memories)
    return sorted(unique_memories, key=lambda m: m.get('relevanceScore', 0), reverse=True)[:10]

undefined

**Update Memory Record**:
```python

**Delete Memory Records**:
```python

---

undefined

job = next(j for j in status['jobs'] if j['jobId'] == job_id)

if job['status'] == 'COMPLETED':
    print(f"Extracted {job['recordsCreated']} new memories")
    break
elif job['status'] == 'FAILED':
    print(f"Extraction failed: {job['error']}")
    break

time.sleep(30)

**Custom Extraction Instructions**:
```python

            1. USER PREFERENCES (high priority):
               - Product preferences
               - Communication style preferences
               - Time/schedule preferences

            2. IMPORTANT FACTS (high priority):
               - Allergies or restrictions
               - Account/membership status
               - Key dates (birthdays, anniversaries)

            3. INTERACTION HISTORY (medium priority):
               - Products purchased
               - Issues resolved
               - Feedback given

            4. CONTEXT HINTS (low priority):
               - Mentioned family members
               - Hobbies or interests
               - Location information

            DO NOT extract:
            - Temporary session-specific details
            - Sensitive financial information
            - Health information beyond allergies
            '''
        }
    }
}

---

from bedrock_agentcore import BedrockAgentCoreApp
from strands import Agent

app = BedrockAgentCoreApp()
memory_client = boto3.client('bedrock-agentcore')

MEMORY_ID = 'memory-xxx'

@app.entrypoint
def invoke(payload):
    user_id = payload.get('user_id')
    user_message = payload.get('prompt')
    session_id = payload.get('session_id', str(uuid.uuid4()))

    # 1. Retrieve relevant memories
    memories = get_relevant_memories(user_id, user_message)
    memory_context = format_memories_for_context(memories)

    # 2. Build enhanced prompt with memories
    enhanced_prompt = f"""
You are a helpful assistant with knowledge about this user.

USER HISTORY AND PREFERENCES:
{memory_context}

CURRENT REQUEST:
{user_message}

Respond helpfully, incorporating relevant knowledge about the user.
"""

    # 3. Run agent
    agent = Agent(model="anthropic.claude-sonnet-4-20250514-v1:0")
    result = agent(enhanced_prompt)

    # 4. Store interaction for future learning
    store_interaction(user_id, session_id, user_message, result.message)

    return {"response": result.message}

def get_relevant_memories(user_id, query):
    """Retrieve relevant memories for context"""
    try:
        response = memory_client.retrieve_memory_records(
            memoryId=MEMORY_ID,
            actorId=user_id,
            retrievalQuery={
                'semanticQuery': query,
                'maxRecords': 5
            }
        )
        return response['memoryRecords']
    except Exception:
        return []

def format_memories_for_context(memories):
    """Format memories as context string"""
    if not memories:
        return "No prior interaction history available."

    lines = []
    for m in memories:
        lines.append(f"- {m['content']}")
    return "\n".join(lines)

def store_interaction(user_id, session_id, user_msg, assistant_msg):
    """Store interaction for memory extraction"""
    memory_client.create_event(
        memoryId=MEMORY_ID,
        actorId=user_id,
        sessionId=session_id,
        event={
            'eventTime': datetime.datetime.now(datetime.timezone.utc).isoformat(),
            'traceId': str(uuid.uuid4()),
            'userMessage': {'content': user_msg}
        }
    )
    memory_client.create_event(
        memoryId=MEMORY_ID,
        actorId=user_id,
        sessionId=session_id,
        event={
            'eventTime': datetime.datetime.now(datetime.timezone.utc).isoformat(),
            'traceId': str(uuid.uuid4()),
            'assistantMessage': {'content': assistant_msg}
        }
    )

undefined

undefined

undefined

undefined

undefined

---