Multi-Agent Shared Memory

Agents can share state via shared memory blocks. This allows agents to have a “shared memory”. You can shared blocks between agents by attaching the same block ID to multiple agents.

In the example code below, we create a shared memory block and attach it to a supervisor agent and a worker agent. Because the memory block is shared, when one agent writes to it, the other agent can read the updates immediately.

1 # install letta_client with `pip install letta-client`
2 from letta_client import Letta
3 
4 # create a client to connect to Letta
5 client = Letta(token="LETTA_API_KEY")
6 
7 # create a shared memory block
8 shared_block = client.blocks.create(
9     label="organization",
10     description="Shared information between all agents within the organization.",
11     value="Nothing here yet, we should update this over time."
12 )
13 
14 # create a supervisor agent
15 supervisor_agent = client.agents.create(
16     model="anthropic/claude-3-5-sonnet-20241022",
17     embedding="openai/text-embedding-3-small",
18     # blocks created for this agent
19     memory_blocks=[{"label": "persona", "value": "I am a supervisor"}],
20     # pre-existing shared block that is "attached" to this agent
21     block_ids=[shared_block.id],
22 )
23 
24 # create a worker agent
25 worker_agent = client.agents.create(
26     model="anthropic/claude-3-5-sonnet-20241022",
27     embedding="openai/text-embedding-3-small",
28     # blocks created for this agent
29     memory_blocks=[{"label": "persona", "value": "I am a worker"}],
30     # pre-existing shared block that is "attached" to this agent
31     block_ids=[shared_block.id],
32 )

Memory blocks can also be accessed by other agents, even if not shared. For example, worker agents can write the output of their task to a memory block, which is then read by a supervisor agent. To access the memory blocks of other agents, you can simply use the SDK clients or API to access specific agent’s memory blocks (using the core memory routes).