Memory, Persistence & Time Travel

1. Memory Taxonomies in Autonomous Agents

An autonomous agent requires three distinct tiers of memory to operate across long horizons:

2. LangGraph Checkpointers: How State is Saved

In LangGraph, a Checkpointer automatically captures a complete StateSnapshotat the end of every node execution (super-step):

• MemorySaver: In-memory dictionary checkpointer. Perfect for unit testing and local development.
• SqliteSaver: Persists checkpoints to a local SQLite database file on disk.
• PostgresSaver: Production-grade enterprise checkpointer. Handles high concurrency, row-level locks, and multi-node clusters.

from langgraph.checkpoint.sqlite import SqliteSaver

# Initialize a persistent SQLite checkpointer
memory = SqliteSaver.from_conn_string("agent_state.db")

# Compile graph with persistence attached
app = workflow.compile(checkpointer=memory)

3. Multi-User Session Isolation via thread_id

When invoking a compiled graph, pass a thread_id in the configuration. LangGraph automatically isolates state so different users never leak context into each other:

# Invoking for User A (thread_101)
config_a = {"configurable": {"thread_id": "thread_101"}}
response_a = app.invoke({"messages": ["Hello, my name is Alice."]}, config=config_a)

# Invoking for User B (thread_102)
config_b = {"configurable": {"thread_id": "thread_102"}}
response_b = app.invoke({"messages": ["Hello, my name is Bob."]}, config=config_b)

# Resume conversation with User A
response_a_2 = app.invoke({"messages": ["What is my name?"]}, config=config_a)
# Model correctly answers: "Your name is Alice."

4. Time Travel: Rewinding and Forking Agent State

One of the most powerful features of LangGraph is Time Travel. Because every state snapshot is immutably stored, you can inspect historical steps, rewind the agent to a previous point in time, and fork execution with modified state:

# 1. Inspect state history for a specific thread
state_history = list(app.get_state_history(config_a))

print(f"Total historical checkpoints recorded: {len(state_history)}")
for snapshot in state_history[:3]:
    print(f"Checkpoint ID: {snapshot.config['configurable']['checkpoint_id']}")
    print(f"Next Node to run: {snapshot.next}")
    print(f"Messages count: {len(snapshot.values['messages'])}\n")

# 2. Rollback & Fork: Select an earlier checkpoint ID
target_checkpoint = state_history[-2].config

# 3. Update or overwrite state at that exact point in time
app.update_state(
    target_checkpoint,
    {"messages": ["Actually, my name is Dr. Alice Walker."]}
)

# 4. Resume execution from the fork!
resumed_response = app.invoke(None, config=target_checkpoint)