AMAZON BEDROCK / SERVICE DETAIL S03 MEMORY

AGENTCORE MEMORYPRACTICAL

CHAPTER 10 / 36
TOPIC create a memory, write events, retrieve
REV 2026.07

New to these words? Sheet 01-A "Key terms" defines session, namespace, and tool. Concept: chapter 09 "Memory".

IN PLAIN WORDS

You will create a memory, write a few chat messages into it, then read back both the raw messages and the facts the agent learned from them. Terms: strategy = the rule for what to remember (SEMANTIC keeps plain facts); actor id = the user; session id = the conversation; namespace = a folder-like path where memories are filed; top_k = how many results to return.

[ 1 ] INSTALL AND CREATE A PROJECT

$ pip install bedrock-agentcore $ npm install -g @aws/agentcore $ agentcore create --name agentcore-memory-quickstart --no-agent $ cd agentcore-memory-quickstart $ python -m venv .venv && source .venv/bin/activate

~~[x]~~ AWS account (aws configure)

~~[x]~~ Python 3.10+

~~[x]~~ Node.js 18+ (for the CLI)

[ 2 ] STEP 1, CREATE A MEMORY WITH A SEMANTIC STRATEGY

$ agentcore add memory --name CustomerSupportSemantic --strategies SEMANTIC $ agentcore deploy # takes 2 to 3 minutes $ agentcore status # verify the memory was created

A memory accepts short-term events by default. A strategy (here SEMANTIC) is what turns them into long-term records.

[ 3 ] STEP 2 AND 3, WRITE EVENTS, THEN RETRIEVE (AgentCore SDK)

# write conversation turns (short-term)
import boto3
from bedrock_agentcore.memory import MemorySessionManager
from bedrock_agentcore.memory.constants import (
    ConversationalMessage, MessageRole)

ctrl = boto3.client('bedrock-agentcore-control',
                   region_name='us-west-2')
memory_id = ctrl.list_memories()['memories'][0]['id']

mgr = MemorySessionManager(memory_id=memory_id,
                            region_name="us-west-2")
session = mgr.create_memory_session(
    actor_id="User1", session_id="OrderSupportSession1")

session.add_turns(messages=[ConversationalMessage(
    "How can I help you today?", MessageRole.ASSISTANT)])
session.add_turns(messages=[ConversationalMessage(
    "My order #35476 has not arrived.", MessageRole.USER)])

turns = session.get_last_k_turns(k=5)        # read short-term

# retrieve from long-term memory

# all records under a namespace path
records = session.list_long_term_memory_records(
    namespace_path="/")

# or a semantic search
records = session.search_long_term_memories(
    query="summarize the support issue",
    namespace_path="/",
    top_k=3,
)

for r in records:
    print(r)

Under the hood these call CreateEvent (short-term) and RetrieveMemoryRecords (semantic, cosine relevance). Long-term extraction runs asynchronously, so allow a short delay before records appear.

[ 4 ] WIRE MEMORY INTO A STRANDS AGENT, AND CLEAN UP

$ agentcore add memory --name MyMemory \ --strategies SEMANTIC,SUMMARIZATION $ agentcore deploy # cleanup when done: $ agentcore remove memory --name CustomerSupportSemantic $ agentcore deploy

# app/MyAgent/memory/session.py (Strands integration)
from bedrock_agentcore.memory.integrations.strands.session_manager \
    import AgentCoreMemorySessionManager
from bedrock_agentcore.memory.integrations.strands.config \
    import AgentCoreMemoryConfig, RetrievalConfig

# env var MEMORY_MYMEMORY_ID is injected at runtime
# pass the manager to your agent in main.py:
#   Agent(model=..., session_manager=get_memory_session_manager(...))

SRC: AWS AgentCore Developer Guide (get started with AgentCore Memory, retrieve records)

DRAWN 2026.06, REV 2026.07.26