This article is part of a series deep-diving into real gritty tech behind Kin. You can find the original article here.

When researching memory systems for AI agents, you’ll find numerous options on the market. Some are relatively simple, built around semantic graphs, while others are more document-oriented. Notable solutions like ZEP with its temporal graphs and Memento, Mem0 offer impressive capabilities. However, after careful evaluation, we determined that building our own memory system was necessary for several compelling reasons.

I'm going to cover them now - but you can also hear more about them in this accompanying video:

Why build a new memory for an AI companion?

Most existing memory systems are developed in Python and designed for server-side deployment. They typically require complex databases like Neo4j or PostgreSQL, which creates a fundamental mismatch with our on-device (edge) architecture. Our solution needs to run entirely on user devices, prioritizing privacy while working within hardware constraints. This technical limitation alone necessitates a different approach with embeddable databases and optimized processing.

Beyond these technical constraints, we discovered that existing memory solutions lack several crucial elements for our use case. Many are heavily document-retrieval oriented, rather than focused on user understanding and human-like memory simulation. We don’t just want to retrieve documents; we want to understand users and simulate human memory processes authentically.

While knowledge graphs (a simple way of visually showing the relationships between information )provide a solid foundation, we found them too limited - especially when confined to triple structures (subject-predicate-object). Modern language models can form much more complex relations with multi-arity entities that require more sophisticated graph structures. Hypergraphs and metagraphs theoretically offer better solutions here, but they lack practical database implementations and sufficient research for our immediate needs. Therefore, we decided to build something using directed graphs in innovative ways, which allows us to model semantic schema packages and multi-entity facts effectively.

Perhaps most importantly, we also needed a system with robust temporal awareness. Human memory doesn’t just store facts - it places them within temporal contexts, with each piece of information having validity periods and connections to specific times. Off-the-shelf AI memory systems like ZEP offer interesting temporal capabilities, but ZEP requires concrete timestamps for all information. In reality, people often talk about time in abstract or incomplete ways (“sometime last summer” versus precise timestamps), and a good memory system should accommodate this flexibility with partial or abstract temporal references.

So, our work was cut out for us.

What kind of memory are we actually building for personal AI?

Our approach directly addresses significant limitations we’ve identified in classical knowledge graphs - by replacing them. Instead, we’ve created a concept graph enhanced with event-based structures that maintain comprehensive temporal awareness throughout the system.

This innovative architecture allows us to model complex semantic facts and schemas with much greater flexibility than traditional knowledge graph approaches. Rather than being limited to simple triples, our system can represent the rich, multi-dimensional relationships that characterize human understanding. Our bipartite graph incorporates complex subgraphs as nodes, representing facts, events, and their interconnections in ways that better mirror human cognition.

A distinguishing feature of our system is that events and life experiences are treated as first-class citizens, equal in importance to facts. Events aren’t merely timestamps attached to data points - they’re fundamental organizing structures within the memory system. This approach enables us to build meaningful relationships between experiences, including crucial causality connections that are typically overlooked in other memory systems.

The integration of semantic meaning with temporal context creates what might be called “semantic spacetime” - understanding entities not just by what they are, but by when they exist, how we learned about them, and what influences they have had on other information. This approach goes beyond simple temporal graphs by focusing on the meaning of entities within temporal contexts and understanding the causal relationships that brought them into existence.

Our system also includes advanced capabilities for smart recall mechanisms and flexible forgetting. Memory isn’t just about remembering everything — it’s equally about forgetting strategically and maintaining appropriate focus. We’ve designed systems that can evaluate which memories deserve attention, which should be readily accessible, and which can be safely deprioritized, mirroring human memory’s selective nature without inheriting its flaws like false memories or unintentional modifications.

Why Kin's memory is uniquely better for an AI chatbot

Our implementation philosophy centers around a hybrid neuro-symbolic approach that combines the strengths of classical AI methods with modern deep learning. The neural component leverages language models and machine learning for pattern recognition and generalization, while the symbolic side employs reasoning on knowledge graphs and logical systems for structured representation and inference.

This balanced combination provides several distinctive advantages over purely neural or purely symbolic approaches. The reasoning and logical systems built atop our knowledge graphs significantly enhance LLM capabilities while simultaneously reducing hallucinations. By grounding language model outputs in a structured knowledge representation, we create a more reliable and consistent agent experience.

Our comprehensive extraction and processing pipeline gives agents the ability to think more deeply about information, going beyond surface-level pattern matching to understand the underlying relationships between concepts, events, and experiences. This depth of processing is essential for creating truly intelligent agents rather than simple retrieval systems.

A critical aspect of our approach is balancing different aspects of memory function. We focus not just on relevance (finding appropriate information) but also on importance (determining what matters most) and attention (deciding where to focus cognitive resources). This multi-dimensional approach to memory prioritization creates a more human-like experience while maintaining computational efficiency.

We’ve also worked to implement sophisticated mechanisms for both recall and forgetting.

While accurate recall is obviously essential, the ability to strategically forget or deprioritize information is equally important for an efficient memory system. Our approach aims to mimic human memory’s selectivity without inheriting its flaws like false memories or unintentional modifications.

The promise degree concept in our system further allows us to build complex networks of agents, extending the memory capabilities beyond individual agents to create collaborative intelligence. This foundation for multi-agent systems opens possibilities for more sophisticated and resilient AI ecosystems in the future.

By taking this comprehensive approach to memory system design, we’re creating a foundation that will enable our agents to understand users more deeply, reason more effectively about complex situations, and provide more helpful and contextually appropriate assistance than would be possible with existing off-the-shelf solutions.