Why We Chose Smrti Over Mem0: The Deep Bet Behind Our AI Companion
The article discusses the limitations of current AI companion products that rely on retrieval-augmented generation (RAG) architectures, and introduces Smrti, an open-source project that takes a different approach to building long-term AI companions.
Why it matters
This article highlights a critical limitation in the current state of AI companions and proposes an alternative approach that could lead to more meaningful and lasting relationships between humans and AI.
Key Points
- 1Current AI companions suffer from a 'structural death spiral' where personality is just a prompt-layer wrapper, and every base-model swap kills the user's 'partner'
- 2RAG-based companions are excellent librarians but do not truly understand companionship, which is about how the user has changed the AI, not just what the user has said
- 3Smrti, an open-source project, takes a more holistic approach to AI memory and personality, incorporating Bayesian truth maintenance, semantic networks, and other techniques beyond simple vector search
Details
The article argues that current AI companion products, which rely on retrieval-augmented generation (RAG) architectures, fundamentally misunderstand what companionship is. RAG-based systems can recall what the user has said in the past, but they do not truly 'remember' or understand how the user has changed the AI's personality over time. The article introduces Smrti, an open-source project that takes a more holistic approach, incorporating Bayesian truth maintenance, semantic networks, and other techniques beyond simple vector search. Smrti aims to build AI companions whose personalities evolve and change based on their interactions with users, rather than just retrieving past conversations. The article suggests that this deeper approach to AI memory and personality is necessary to create truly long-lasting and meaningful AI companions.
No comments yet
Be the first to comment