langgraph-agentram

An AgentRAM-backed BaseStore for LangGraph. It gives your LangGraph agents cross-thread long-term memory through a hosted key-value API, with no vector database and no embedding pipeline to run.

LangGraph's long-term memory is built on stores: JSON values organized by a namespace and a key. AgentRAM is a hosted memory API with exactly that shape, so this adapter is a thin, honest bridge between the two.

Install

pip install langgraph-agentram

Get a free AgentRAM API key at agentram.dev. New accounts start with 1,000 credits, no card required.

Use it

Pass the store to your agent and it gains memory that survives across threads:

from langchain.agents import create_agent
from langgraph_agentram import AgentRAMStore

store = AgentRAMStore(api_key="agentram_your_key_here")

agent = create_agent("claude-sonnet-4-6", tools=[], store=store)

Or use the store directly:

store.put(("memories", "user-1"), "language", {"value": "French"})
item = store.get(("memories", "user-1"), "language")
print(item.value)  # {"value": "French"}

# list a namespace, or text-search within it
store.search(("memories", "user-1"))
store.search(("memories", "user-1"), query="French")

Memory tools for any LangChain agent

The store plugs into LangGraph's native memory slot. If you would rather give an agent memory it can call as tools, and have it work in plain LangChain too, use the tools:

from langgraph_agentram import create_agentram_tools

tools = create_agentram_tools(api_key="agentram_your_key_here")

from langchain.agents import create_agent
agent = create_agent("claude-sonnet-4-6", tools=tools)

create_agentram_tools returns four tools the agent can call: save_memory, recall_memory, search_memory, and list_memories. They are backed by the same AgentRAM store, so the same honest limits apply. Give separate agents separate namespaces to keep their memories apart:

tools = create_agentram_tools(api_key="agentram_...", namespace=("memories", "user-1"))

How it maps to AgentRAM

• A namespace tuple becomes an AgentRAM agent_id, joined with / (for example ("memories", "user-1") becomes memories/user-1).
• The key is the AgentRAM key.
• The value dict is JSON-encoded into AgentRAM's value field.

Honest limits

This adapter does what AgentRAM does, and nothing it does not.

• search is a text match, not semantic ranking. score is always None. That is the point: memory without a vector database.
• search treats the namespace you pass as a full namespace, not a prefix to walk into nested sub-namespaces.
• list_namespaces is not supported. AgentRAM has no endpoint to enumerate namespaces, so the adapter raises rather than return a wrong answer. Track namespaces in your own app if you need them.
• A namespace maps to an agent_id capped at 100 characters, and a value is capped at 5,000 characters. Both raise a clear error if exceeded, rather than truncating.
• Per-item TTL is not mapped in this version.

Config

• api_key (required): your key, starts with agentram_.
• base_url (optional): defaults to https://api.agentram.dev.

Links

• AgentRAM: https://agentram.dev
• API docs: https://agentram.dev/docs.html
• Issues: https://github.com/seanmarkwei/langgraph-agentram/issues

MIT