There’s a quiet shift underway in how we think about AI, and it begins with memory.

For decades, infrastructure has been about speed, storage, and scale. But as AI systems become more agentic, more conversational, and more embedded into everyday workflows, a deeper limitation has surfaced: they forget. Every interaction begins from scratch, every context rebuilt, every nuance at risk of being lost.

It’s a problem Prashant Malik has been thinking about for a long time.

Best known as the co-creator of Apache Cassandra and among the earliest engineers at Facebook, Malik has spent over two decades building and advising on distributed systems that quietly power the internet. With CortexDB, he’s now turning that experience toward a new frontier, AI infrastructure.

Rethinking Memory, From First Principles

At its core, CortexDB is designed as a long-term memory layer for AI systems. But what makes it compelling is not just what it does, it’s how differently it approaches the problem.

Today’s AI agents, despite their sophistication, are fundamentally stateless. Memory, where it exists, is often stitched together using large language models that continuously rewrite and compress data. It’s an approach that feels clever until you consider the cost.

Each rewrite risks losing information. Each summary is a decision, what stays, what goes. And once it’s gone, it’s gone for good.

CortexDB challenges this paradigm entirely.

Instead of rewriting memory, it preserves it.

Every piece of data is appended to an immutable event log, never overwritten, never reduced to a fragile summary. A lightweight model works in the background to extract entities and relationships, but the original data remains intact, always available, always re-queryable. It’s a subtle but profound shift: memory is no longer something AI approximates, but something it retains.

Building Context That Evolves Over Time

From this continuous stream of events, CortexDB constructs a temporal knowledge graph, mapping not just entities and relationships, but also causality and provenance. It’s memory that evolves, not erases.

When an AI agent needs context, CortexDB doesn’t rely on a single retrieval method. Instead, it assembles information through a hybrid approach, vector search, full-text matching, graph traversal, and adaptive ranking, bringing together exactly what’s needed, when it’s needed.

The result isn’t just better recall. It’s a fundamentally different way of understanding context.

In controlled benchmarks across production-scale scenarios, this approach has shown a significant performance gap compared to conventional systems. Not because of tuning or optimization, but because of structure, because you can’t retrieve what you’ve already discarded.

Scaling, The Cassandra Way

If memory is the heart of CortexDB, scalability is its backbone.

Drawing from the design philosophy behind Apache Cassandra, the system is built to scale horizontally from the ground up. Through consistent hashing, partition-aware data placement, and leaderless replication, every component, whether it’s an index, a graph shard, or a vector store, is distributed by design.

There’s no transition from prototype to scale. No architectural rewrite waiting in the future.

A single-node deployment runs the same code as a hundred-node cluster. Adding capacity is as simple as adding another node, with the system rebalancing itself automatically and without downtime. It’s infrastructure that grows the way the internet itself does, organically, seamlessly, and without friction.

A New Layer in the AI Stack

What CortexDB proposes isn’t an incremental improvement. It’s a redefinition.

For years, the focus in AI has been on models, their size, their accuracy, their capabilities. But as applications mature, it’s becoming clear that models alone are not enough. They need memory that is reliable, scalable, and lossless.

CortexDB positions itself as that missing layer.

A system where memory isn’t rewritten, but remembered. Where scale isn’t an afterthought, but a foundation. And where AI systems can finally move beyond stateless interactions toward something closer to continuity.

It’s a shift that feels both technical and philosophical.

Because in the end, intelligence isn’t just about generating answers. It’s about remembering what mattered.