Long Running Agents

A Pydantic AI agent that remembers across sessions, can create its own tools, and delegates to specialists. Memory, sandbox, and tasks in one stack.

Source

Why this exists

Problem: Most agents forget between runs. They can't recall past conversations, and they can't extend their own tool set.

Solution: Long Running Agents gives you persistent memory (SQLite + ChromaDB), cross-session recall, and dynamic tool generation—so your agent remembers, learns, and adapts over time.

Audience: Developers building long-lived, memory-aware agents with Pydantic AI.

What makes it different

Feature	What it does	Why it stands out
Cross-session memory	get_recent_conversations(all_sessions=True) + search_memory	Most agents forget between runs. This one recalls past turns across sessions.
Dynamic tool generation	Agent creates new tools at runtime via generate_tool	Extends its own tool set with AST validation and persistence.
Subagent delegation	Code and research specialists	Routes work to focused subagents instead of one monolithic agent.
Hybrid retrieval	Vector + keyword module	Optional hybrid search (semantic + keyword) available in the memory layer.
Pydantic AI	Typed agent framework	Uses Pydantic AI instead of LangChain/LlamaIndex.
Sandbox + memory + tasks	All in one stack	Memory, code execution, and task tracking in a single package.

What it is (and isn't)

• Is: A chat-driven AI agent with persistent memory, code execution, and task tracking. You interact via a terminal loop; the agent responds using its tools; context is saved across runs.
• Isn't: A workflow automation engine. No schedules, triggers, or DAGs. It's an interactive assistant, not Zapier or Airflow.

Install

Prerequisites: Python 3.10+, OpenAI API key

# From PyPI
pip install long-running-agents

# From GitHub
pip install git+https://github.com/prith27/lra.git

# From source (clone first)
git clone https://github.com/prith27/lra.git
cd lra
pip install -e .

Setup

Set your OpenAI API key (required). Choose one:

# Option A: Export in shell
export OPENAI_API_KEY=sk-your-key-here

# Option B: Create .env file in your project directory
echo "OPENAI_API_KEY=sk-your-key-here" > .env

# Option C: Use lra init to create .env template, then add your key
lra init

Get a key at platform.openai.com/api-keys.

Quick start

lra chat

Note: Basic chat and memory work without the sandbox. For code execution or dynamic tool creation, start the sandbox first in a separate terminal (see Sandbox).

How it works

1. Run lra chat to start the chat loop.
2. Type a message; the agent may call tools (search memory, run code, create tasks, delegate to subagents).
3. Each turn is persisted to SQLite and ChromaDB so the agent can recall past context in future runs.
4. Each run gets a new session ID, but get_recent_conversations(all_sessions=True) and search_memory allow cross-session recall.

Agent tools

Tool	Purpose
search_memory	Semantic search over past summaries and facts (ChromaDB)
write_memory	Store facts or summaries for later recall
get_recent_conversations	Fetch recent turns (session or all sessions)
create_task, update_task_status, list_tasks	Track multi-step work
create_sandbox, execute_code	Run Python in isolated containers
delegate_code_task, delegate_research_task	Hand off to specialist subagents
generate_tool	Create new tools at runtime when no existing tool fits

Library usage

import asyncio
from long_running_agents import run_agent, AgentDeps, StructuredMemoryStore, VectorMemoryStore
from config import SANDBOX_URL

async def main():
    structured = StructuredMemoryStore()
    vector = VectorMemoryStore()
    await structured.init_db()

    deps = AgentDeps(
        session_id="my-session",
        structured_store=structured,
        vector_store=vector,
        sandbox_base_url=SANDBOX_URL,
    )

    output, messages = await run_agent("What can you do?", deps)
    print(output)
    await structured.close()

asyncio.run(main())

Examples

Examples are included in the package. After installing:

python -m long_running_agents.examples.01_basic_chat
python -m long_running_agents.examples.02_single_turn "Your question here"

Example	Description
01_basic_chat	Chat loop: multiple turns, memory persists across runs
02_single_turn	One-off query: ask a question, get a response, exit

When installed from source, see long_running_agents/examples/README.md for details.

CLI

lra init              # Create .env and show setup instructions
lra chat              # Start the agent chat loop
lra list-tools        # List static and dynamic tools
lra inspect-tool X    # Inspect a dynamic tool
lra list-memory -s SESSION  # List memory for a session

Framework commands

Create and run custom agents with their own system prompts:

lra create-agent [name]     # Create agent dir (default: my_agent). Use --prompt or enter interactively
lra run [path]              # Run a custom agent (path to agent dir or main.py)
lra list-agents             # List agent directories
lra config                 # Show config

Create and manage tools:

lra create-tool             # Create a tool interactively
lra create-tool --file X    # Create a tool from a Python file
lra export-tools [-o path]  # Export dynamic tools to static file
lra validate-tool FILE      # Validate a tool file in sandbox

Note: my_agent/ and *_agent/ are in .gitignore by default so user-created agents are not committed. Add your own pattern to .gitignore if you want to ignore different agent dirs.

Configuration

Variable	Description	Default
OPENAI_API_KEY	OpenAI API key	(required)
SANDBOX_URL	Sandbox API base URL	http://localhost:8000
DATABASE_URL	SQLAlchemy async URL	sqlite+aiosqlite:///./data/agent_memory.db
VECTOR_STORE_PATH	ChromaDB path	./data/chroma_db

Sandbox (optional)

The sandbox enables code execution and tool validation. It includes requests and httpx for HTTP-fetching tools. Basic chat and memory work without it.

Prerequisites: Docker must be installed and running. On macOS, open Docker Desktop and wait until it's ready before starting the sandbox. The sandbox spawns isolated containers for code execution.

Run order

1. Start Docker (e.g. open Docker Desktop on Mac).
2. Start the sandbox in one terminal.
3. Run the agent in another terminal.

# Terminal 1: Ensure Docker is running, then start sandbox (keep running)
python -m uvicorn sandbox.server:app --reload --port 8000

# Terminal 2: Run agent
lra chat

Options

Option	Command	When to use
A: Local	python -m uvicorn sandbox.server:app --reload --port 8000	Development; run from project root with deps installed
B: Docker Compose	docker compose up sandbox	Fully containerized; no local Python needed for sandbox

First run

On first start, the kernel image (longrunningagents-kernel:latest) is built automatically. This may take a minute.

Rebuild kernel

If you updated sandbox/Dockerfile (e.g. added packages), rebuild the kernel:

docker rmi longrunningagents-kernel:latest
# Then restart the sandbox

Project structure

├── agents/              # Main agent and subagents
├── tools/               # Memory, sandbox, task tools
├── sandbox/             # Sandbox API and kernel
├── memory/              # Structured and vector stores
├── schemas/             # Pydantic models
├── long_running_agents/ # Package exports
├── long_running_agents/examples/  # Example recipes (shipped with package)
├── cli.py               # CLI entry point
├── config.py
├── main.py
└── pyproject.toml

Development

pip install -e ".[dev]"
pytest tests/ -v
mypy agents tools memory schemas

License

MIT