LongMem4LangGraph — Long-term Memory for LangGraph

SQLite-backed persistent memory for LangGraph that gives your agents AGNO-style full conversation history, crash recovery, and queryable state — without Redis, PostgreSQL, or external dependencies.

Why?

LangGraph is great for building agent state machines, but its default memory is in-memory only — lost on server restart, inaccessible outside the graph, and agents only see the current state, not the full history of what came before.

LongMem4LangGraph fixes this with a single SQLite file:

Feature	LangGraph Default	LongMem4LangGraph
Persistence	❌ Lost on restart	✅ SQLite file
AGNO-style history	❌ State only	✅ Full conversation memory
Crash recovery	❌ Manual	✅ Auto-detect + resume
Queryable	❌ Not possible	✅ SQL queries
Cross-module access	❌ Graph only	✅ Any code can read
Concurrent reads	❌ Single thread	✅ WAL mode
Setup	❌ Needs Redis/PG	✅ pip install + ready

Install

pip install longmem4langgraph

Quick Start

1. Persistent Checkpoint Saver

Replace LangGraph's default in-memory checkpointer:

from langgraph.graph import StateGraph
from longmem4langgraph import SqliteSaver

# Create a graph with persistent SQLite state
builder = StateGraph(AgentState)
# ... add nodes, edges ...
graph = builder.compile(checkpointer=SqliteSaver("state.db"))

# State survives restarts — run this tomorrow and it picks up where it left off

2. AGNO-Style Agent Memory

Give each agent access to the full history of what previous agents did:

from longmem4langgraph import HistoryStore

history = HistoryStore("state.db")

def agent2_node(state):
    # Read last 20 turns of agent conversation (like AGNO memory)
    context = history.get_context(state["request_id"], last_n=20)
    
    # Inject into prompt
    prompt = f"Previous analysis:\n{context}\n\nNow generate..."
    ...

3. Crash Recovery

Recover pipelines that were interrupted by a crash:

from longmem4langgraph import recover_pipelines

# On system startup:
failed = await recover_pipelines("state.db", graph)
print(f"Recovered {len(failed)} interrupted pipelines")

4. Query Everything

from longmem4langgraph import StateManager

sm = StateManager("state.db")

# Full pipeline audit
summary = sm.get_pipeline_summary("req-123")
print(f"Node count: {summary['total_nodes']}")
print(f"LLM calls: {summary['total_agent_calls']}")
print(f"Total cost: ${summary['total_cost']:.2f}")

# SQL query for advanced analytics
expensive = sm.query("""
    SELECT agent_name, SUM(cost_cents) as total
    FROM agent_calls GROUP BY agent_name
    ORDER BY total DESC
""")

Module Reference

Module	Class/Function	Purpose
longmem4langgraph	StateManager	SQLite-backed Singleton — all-in-one state, history, audit
longmem4langgraph.saver	SqliteSaver	LangGraph BaseCheckpointSaver implementation
longmem4langgraph.history	HistoryStore	AGNO-style full conversation memory
longmem4langgraph.recovery	recover_pipelines()	Auto-detect and resume crashed pipelines

Example: Full Agent Pipeline

import asyncio
from langgraph.graph import StateGraph, END
from typing import TypedDict
from longmem4langgraph import SqliteSaver, HistoryStore

class AgentState(TypedDict):
    request_id: str
    input_text: str
    agent1_result: str
    agent2_result: str

saver = SqliteSaver("pipeline.db")
history = HistoryStore("pipeline.db")

def agent1(state: AgentState):
    result = f"Agent 1 analyzed: {state['input_text'][:50]}..."
    history.add(state["request_id"], "agent1", result)
    return {"agent1_result": result}

def agent2(state: AgentState):
    context = history.get_context(state["request_id"], last_n=5)
    result = f"Agent 2 used context: {context[:100]}..."
    history.add(state["request_id"], "agent2", result)
    return {"agent2_result": result}

builder = StateGraph(AgentState)
builder.add_node("agent1", agent1)
builder.add_node("agent2", agent2)
builder.set_entry_point("agent1")
builder.add_edge("agent1", "agent2")
builder.add_edge("agent2", END)

graph = builder.compile(checkpointer=saver)

# Run
config = {"configurable": {"thread_id": "test-1"}}
result = graph.invoke(
    {"request_id": "test-1", "input_text": "Hello agents!"},
    config
)
print(result)

License

MIT — free for everyone. Use it, improve it, share it.