Memobase

Memobase is a user profile-based memory system designed to bring long-term user memory to your LLM applications. Whether you're building virtual companions, educational tools, or personalized assistants, Memobase empowers your AI to remember, understand, and evolve with your users.

Memobase offers the perfect balance for your product among various memory solutions. At Memobase, we focus on three key metrics simultaneously:

• Performance: Although Memobase is not specifically designed for RAG/search tasks, it still achieves top-tier search performance in the LOCOMO benchmark.
• LLM Cost: Memobase includes a built-in buffer for each user to batch-process their chats, allowing the overhead to be distributed efficiently. Additionally, we carefully design our prompts and workflows, ensuring there are no "agents" in the system that could lead to excessive costs.
• Latency: Memobase works similarly to the memory system behind ChatGPT: for each user, there is always a user profile and event timeline available. This allows you to access the most important memories of a user without any pre-processing, but only few SQL operations, keeping online latency under 100ms.

Check out the profile result (compared with mem0) from a 900-turns real-world chatting:

{
  "basic_info": {
    "language_spoken": ["English", "Korean"],
    "name": "오*영"
  },
  "demographics": {
    "marital_status": "married"
  },
  "education": {
    "notes": "Had an English teacher who emphasized capitalization rules during school days",
    "major": "국어국문학과 (Korean Language and Literature)"
  },
  "interest": {
    "games": "User is interested in Cyberpunk 2077 and wants to create a game better than it",
    "youtube_channels": "Kurzgesagt",
    ...
  },
  "psychological": {...},
  "work": {"working_industry": ..., "title": ..., },
  ...
}

🎉 Recent Updates

• 0.0.40: we updated the internal workflows in Memobase, reducing the number of LLM calls in a single run from approximately 3-10 times to a fixed 3 times, which reduces token costs by approximately 40-50%. (Consider updating your Memobase version!)
• 0.0.37: we added fine-grained event gist, enabling the detailed search on users' timeline. Re-ran the LOCOMO benchmark and we're SOTA!
• 0.0.36: we updated the search of context api, making the search take between 500~1000ms (depending on the embedding API you're using). Also, you can pass a prompt template to the context api to pack memories directly into prompt.

📖 Table of Contents

• Table of Contents
• Core Features
• Get Started
• Step-by-step breakdown
  • 1. Make sure you're connected
  • 2. Manage Users
  • 3. Insert Data
  • 4. Get your Memory
  • 5. Integrate memory into your prompt
• What's next?
• Why/Where should I use Memobase?
  • Remember the users
  • User analysis and tracking
  • Sell something to your customers.
• Documentation
• Stay Updated
• Support
• Contribute
• License

Core Features

🎯 Memory for User, not Agent

Define and control exactly what user information your AI captures.

📈 SOTA

Check out performance on public benchmark against mem0, langmem, zep...

📅 Time-aware Memory

Memobase has more than user profiles, it also records user event. User event is essential to answer time-related question, see how we can improve temporal memory much better than other memory solutions.

🖼️ Controllable Memory

Among all types of memory, only some may enhance your product experience. Memobase offers a flexible configuration for you to design the profile.

🔌 Easy Integration

Minimal code changes to integrate with your existing LLM stack with API, Python/Node/Go SDK.

⚡️ Batch-Process:

Memobase offers every user a buffer to batch processing the chats after the conversation. Fast & Cheap.

🚀 Production Ready

Memobase is building with FastAPI, Postgres and Redis, supporting request caching, authing, telemetry... Fully dockerized.

How Memobase works?

Get Started

https://github.com/user-attachments/assets/eb2eea30-48bc-4714-9706-e417ae1931df

1. Start your Memobase server locally. If you don't want to be bothered, Memobase Cloud provides a free tier enough for your testing
2. You should have the below two things to continue:
   1. A project url. (local: http://localhost:8019 , cloud https://api.memobase.dev)
   2. A project token. (local: secret , cloud sk-proj-xxxxxx)
3. Install the Python SDK: pip install memobase
4. Below tutorial is for Python User. For other language and API, check this.

Step-by-step breakdown

1. Make sure you're connected

from memobase import MemoBaseClient, ChatBlob

client = MemoBaseClient(
    project_url=PROJECT_URL,
    api_key=PROJECT_TOKEN,
)
assert client.ping()

2. Manage Users

uid = client.add_user({"any_key": "any_value"})
client.update_user(uid, {"any_key": "any_value2"})
u = client.get_user(uid)
print(u)

# client.delete_user(uid)

3. Insert Data

In Memobase, all types of data are blobs for a user, which can be inserted, retrieved, and deleted:

messages = [
  {
      "role": "user",
      "content": "Hello, I'm Gus",
  },
  {
      "role": "assistant",
      "content": "Hi, nice to meet you, Gus!",
  }
]
bid = u.insert(ChatBlob(messages=messages))
print(u.get(bid)) # not found once you flush the memory.

# u.delete(bid)

By default, Memobase will remove the blobs once they're processed. This means that apart from the relevant memory, your data will not be stored with Memobase. You can persist the blobs by adjusting the configuration file.

4. Get your Memory

u.flush(sync=True)

By default, Memobase will flush the buffer asynchronously. You can set sync=True to wait for the buffer to be processed.

And what will you get?

print(u.profile(need_json=True))

# results
{
  "basic_info": {
    "name": {
      "content": "Gus",
      "id": ...,
      "created_at": ...
    }
  }
}

u.profile() will return structured profiles that are learned from this user, including topic, sub_topic and content. As you insert more blobs, the profile will become better.

5. Integrate memory into your prompt

Memobase has a context api to pack everything you need into a simple string, where you can insert it into your prompt directly:

print(u.context(max_token_size=500, prefer_topics=["basic_info"]))

Something like:

# Memory
Unless the user has relevant queries, do not actively mention those memories in the conversation.
## User Background:
- basic_info:name: Gus
...

## Latest Events:
...

Checkout the detail params here.

What's next?

• Run script: Checkout the quickstart script for more details
• Design your profile!: You may want to explore the customization of Memobase to make sure the system works as your expectation.
• Full-stack Chatbot with Memobase: Check Memobase-Playground. An open-source, full-stack template AI Chatbot with long-term memory. live-demo
• Web UI for Memobase: Check Memobase-Inspector. An open-source UI for your Memobase project with user table, usage chart and test playground. live-demo

Why/Where should I use Memobase?

Remember the users

By placing profiles into your AI (e.g. system prompt).

PROFILES = "\n".join([p.describe for p in u.profile()])

print(PROFILES)
# basic_info: name - Gus
# basic_info: age - 25
# ...
# interest: foods - Mexican cuisine
# psychological: goals - Build something that maybe useful
# ...

User analysis and tracking

Too much information is hidden in the conversations between users and AI, that's why you need a new data tracking method to record user preference and behavior.

PROFILES = u.profile()

def under_age_30(p):
  return p.sub_topic == "age" and int(p.content) < 30

def love_cat(p):
  return p.topic == "interest" and p.sub_topic == "pets" and "cat" in p.content

is_user_under_30 = (
    len([p for p in profiles if under_age_30(p)]) > 0
)
is_user_love_cat = (
  len([p for p in profiles if love_cat(p)]) > 0
)                       
...

Sell something to your customers.

Not everyone is looking for Grammarly, it's always nice to sell something your users might want.

def pick_an_ad(profiles):
  work_titles = [p for p in profiles if p.topic=="work" and p.sub_topic=="title"]
  if not len(work_titles):
    return None
  wt = work_titles[0].content
  if wt == "Software Engineer":
    return "Deep Learning Stuff"
  elif wt == "some job":
    return "some ads"
  ...

Documentation

For detailed usage instructions, visit the documentation.

Stay Updated

Star Memobase on Github to support and receive instant notifications!

Support

Join the community for support and discussions:

• Join our Discord 👻

• Follow us on Twitter 𝕏

Or just email us ❤️

Contribute

• Check out our Changelog first, make sure the feature you want has not been developed or is currently being planned.:)
• Go through Contributing document to setup and contribute to Memobase.

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.