MemGPT

Building persistent LLM agents with long-term memory 📚🦙

[yags @yagilb]

If you’re in NYC next week come hang out with @lmstudio and @Letta_AI (of memGPT fame) teams and learn about stateful local agents

[Charles Packer @charlespacker]

the new frontier: AI agent hosting/serving 👾🛸

the AI/LLM agents stack is a significant departure from the standard LLM stack. the key difference between the two lies in managing state: LLM serving platforms are generally stateless, whereas agent serving platforms need to be stateful (retain the agent state server-side)

[Charles Packer @charlespacker]

thank you @trychroma 🟦🟥🟨 for hosting the 2nd @MemGPT LLM agents meetup at their beautiful mission office

if you're interested in (1) in-person events + (2) LLM agents + (3) AI research.. follow our luma calendar so you don't miss the next one 👉 https://t.co/tnKKtrD5zD

[MemGPT @MemGPT]

✨The second ever @MemGPT LLM agents developer meetup is happening today at @trychroma ! ✨

Expect to learn about agents-as-a-service, agents APIs, persistent agents and giving LLMs long-term memory!

https://t.co/380ma8XiQT

[MemGPT @MemGPT]

In the new MemGPT 0.3.21 release, this is enabled by default for “weaker” models such as gpt-4o, gpt4o-mini, and gpt-3.5-turbo . If you’re finding your MemGPT agents are calling tools but are forgetting to generate internal monologue, try running memgpt run with `--no-content`!

[MemGPT @MemGPT]

💪 We just dropped a major performance update for MemGPT agents using gpt4o-mini 💪

The most cost effective way to run MemGPT agents via OpenAI is now gpt4o-mini - try it out on the latest 0.3.21 release!

[MemGPT @MemGPT]

Previously certain models and model APIs struggled to generate internal monologue to accompany tool calls in MemGPT - you can now get around this issue with the `--no-content` flag in the MemGPT CLI.

[MemGPT @MemGPT]

We look forward to seeing all the new memory modules you create! For information, check out the full 0.3.19 release notes here: https://t.co/oHLT92RNaO

[MemGPT @MemGPT]

Agents will default to using the ChatMemory class, which has the original human/memory fields and memory editing functions in MemGPT:

[MemGPT @MemGPT]

Custom tools and memory classes can now both be specified in the agent creation API:

[MemGPT @MemGPT]

✨ MemGPT now supports custom memory modules! ✨

This allows developers to both define custom memory fields (instead of just human/persona) as well as custom memory editing functions (rather than core_memory_append/replace). 🧵

[Jaya Gupta @JayaGup10]

@MemGPT at the @arizeai conference!

[Happy @123wimi]

🌐 Memory Layer of AI Agents 🌐

Personalization (Memory)🧠

Personalization dynamically adjusts and customizes AI agents' responses and functions based on users' historical behavior, preferences, and specific needs. This enhances user experience, making AI agents more relevant and responsive.

🔹 WhyHowAI @chiajy2000
- Features: Offers personalized recommendations and response optimization.
- Pros: Automatically creates knowledge graphs and integrates with existing workflows, building effective RAG solutions.

🔹 Cognee @cognee_
- Features: Provides personalized services by analyzing user interaction data.

🔹 Graphlit @graphlit
- Features: Uses user data for personalized recommendations.

🔹 LangMem
- Features: Focuses on personalized memory functions, enabling AI agents to remember user preferences and historical interactions.

🔹 MemGPT @MemGPT
- Features: Combines GPT models for personalized response generation.
- Pros: Represents Memory-GPT, a system designed to improve LLM performance through advanced memory management, overcoming fixed context window challenges.

Storage 💾

Storage provides efficient and reliable data storage solutions for agents. These systems must handle large amounts of data and support fast read/write operations to ensure efficient AI model operation.

🔹 Pinecone @pinecone
- Features: High-performance vector database, supports fast data retrieval.

🔹 Chroma @trychroma
- Features: Efficient data storage solutions, open-source vector database designed for AI and embedded applications.

🔹 Weaviate @weaviate_io
- Features: Open-source vector database, supports content-based retrieval and storage.

🔹 MongoDB @MongoDB
- Features: Popular NoSQL database, provides flexible storage and retrieval functions.

Context 🗣���

Context enables AI agents to understand and utilize contextual information in conversations or tasks to provide more accurate and relevant responses. This ensures agents maintain coherence and understand more complex user needs.

🔹 Unstructure @UnstructuredIO
- Features: Open-source project focused on providing robust context management functions, allowing AI agents to understand and utilize context information in conversations or tasks, resulting in more coherent and intelligent responses.

#AI #MemoryLayer #Personalization #Storage #Context #TechStack #Innovation #Startup

[Milvus @milvusio]

🤖 @MemGPT = Easily build and deploy stateful LLM agents!
🗄️ #Milvus = Powerful vector database for your data!
🔥 MemGPT + Milvus = Create agents with strong memory and access to massive-scale external knowledge base.

🔗 https://t.co/CRl6G5o8Py

#AI #RAG #Tech

[MemGPT @MemGPT]

and @milvusio !
https://t.co/k56lQfGxmH

[MemGPT @MemGPT]

@MemGPT now supports @qdrant_engine !

[Qdrant @qdrant_engine]

MemGPT 🧠 - A system to build and deploy stateful LLM agents.

Based on the amazing "MemGPT: Towards LLMs as Operating Systems" paper.
https://t.co/ncVb6oj4ia

[MemGPT @MemGPT]

🚄 token-streaming in the MemGPT REST API

You can now stream tokens back from the MemGPT server when sending an agent messages! This includes streaming of both agent *steps* and tokens within a single step:

[MemGPT @MemGPT]

MemGPT v0.3.18 just landed some big new features:
🔧 tool creation in the MemGPT Python Client
🚄 token-streaming in the MemGPT REST API

[MemGPT @MemGPT]

🔧 tool creation in the MemGPT Python Client

You can now create tools directly inside Python when using the MemGPT Python client!

[MemGPT @MemGPT]

running live demos with @MemGPT agents at @MindsDB 🧠✍️