Cortex v4.0: A Memory Engine That Actually Learns, And +22 Patent Claims to Practice

By Robert Briggs, Founder

When most people say “machine learning” today, they mean a model that was trained once, frozen at inference time, and now answers questions from that frozen weight set. The “learning” happened weeks or months ago in a training run. After deployment, the model doesn't actually learn anything new about you, your project, or your domain. It pattern-matches against what it knew when training stopped.

We've built something different.

The Cortex memory engine inside CxMS (powered by what we call the KMAP protocol) does something the dominant pattern in the industry doesn't do. It refines its own classification accuracy continuously, while it's running, across every dimension that matters for retrieval. Priority. Tagging. Type assignment. Relationship strength. Decay rate. Scope of applicability. Every interaction is a feedback signal, and the system updates its understanding of what each memory means, when it should surface, and how strongly to weight it against alternatives.

That isn't a metaphor for learning. That's literally the definition of machine learning. A system that improves its performance on a task by ingesting more data over time. The industry just doesn't tend to apply the phrase to memory systems, because most memory systems are dumb stores. Ours isn't. It's a learning system that happens to also persist memory as a side effect.

How this got here

Honest version of the origin: I wanted my coding agent to stop forgetting stuff. That's it. Every session I'd lose context I'd already given it. Every refactor I'd watch it walk back into the same wrong choices. It made me crazy.

So the first version was just persistence. Remember what we already decided. Basic.

Then the same kinds of failures kept showing up. Same misreadings, same wrong calls. Storage alone wasn't fixing it because not every memory weighs the same. Repeats need to count more than routine notes. I pulled in human-style emotional weighting for that... let the system bump importance on the things that kept stinging.

That helped. The memory layer kept improving on its own track, getting richer over time.

In parallel, a different insight emerged. Some of what I'd been treating as “memory failures” weren't only about memory. They were about position. Where the memory layer sits in the AI's lifecycle... when it gets consulted... whether it's even visible at the moment of inference. The memory was often there. We just weren't in a spot to use it right. That turned out to be its own architectural problem, and a bigger one than I expected.

Once both tracks were stable enough to build on (memory richness AND positioning), mapping the whole thing onto biology became deliberate. The architecture is grounded in academic neuroscience now. Cortical regions. Hippocampal episodic encoding. The amygdala weighting emotionally significant memories against neutral ones. Consolidation moving short-term memories into stable long-term storage during specific phases. The Cortex engine has correlates for all of those. Memory categories that mirror human memory taxonomy: semantic facts, episodic events, procedural skills, autobiographical identity, spatial knowledge, social relationships, prospective intent. A priority system that surfaces the equivalent of “things you'd never forget” and lets routine details fade gracefully. A consolidation cycle that runs in the background to merge similar memories, strengthen confirmed patterns, and let stale data decay.

Then we improved on the biology in places where digital substrate makes it possible to do better. Human memory is reconstructive. You don't recall a memory, you re-author it each time, which is why eyewitness testimony is unreliable. Our memory system is content-addressed and integrity-hashed... the same query returns the same memory, byte-identical, until something explicitly updates it. Human memory has emotional bias that distorts recall under stress. Our system captures the emotional weighting (because it's useful for retrieval) but separates it from the underlying fact, so the bias is observable and adjustable rather than baked in. Human memory has a single read pathway. Ours can fan out across multiple retrieval modes (exact tag match, similarity-boosted, filtered, or LLM-driven graph traversal) and the system itself learns which mode produces the best results for which query types.

The biology was the blueprint, eventually. The engineering kept being an upgrade.

Built into the agent we're shipping next

We're putting this engine inside our own AI coding agent. Phase one of that agent runs on Anthropic's Claude as the inference layer, with our Cortex engine as the memory layer underneath. Claude is the strongest commercial model available right now for the kind of careful, structured work this agent is designed to do. Multi-file code changes. Long-context reasoning. Tool use. Working alongside a developer rather than in front of one. Pairing it with Cortex means the agent doesn't reset between sessions. It learns the project. It learns the developer's preferences. It learns the corrections that have already been made and doesn't keep walking into the same wrong patterns.

That's the near-term ship. The longer-term plan is our own model trained on our own data. The substrate-independence of the architecture means swapping the inference layer is a configuration change, not a redesign. The same Cortex engine that works behind Claude today will work behind a model we own tomorrow, and the memory accumulated during the Claude phase carries forward.

We're also doing custom development work for teams that want this architecture wired into something other than what we ship by default. Different command-line tools, different model providers, proprietary in-house LLMs. The engine is portable. If you want it inside your own developer workflow on your own model, we'll build it.

From paper to running code: a measurable jump

Patent applications are easy to count. What matters more is how much of what those patents claim is real, working software running in production. That number (reduction to practice) is the honest measure of whether a portfolio is theoretical or operational.

As of v4.0 (April 30, 2026): 65 of 156 filed claims (42%) are reduced to practice in shipping code. That's up from 44 of 156 (28%) before this build. A jump of +21 claims in a single release. On the software side specifically, 66% of all software-governance claims are now running in production, up from 44%. Agent lifecycle management is now at 100% reduction to practice. Three more patents are above 75%: the Cortex memory engine at 95%, cross-vendor consensus at 86%, and the persistent-memory tag-retrieval architecture at 79%.

We track this internally per-claim with code citations against specific functions and methods. When we convert these provisionals to non-provisionals next year, every claim with reduction-to-practice evidence is significantly harder to challenge. An examiner can argue an idea is obvious. They can't argue that running, working code is hypothetical.

The 33 hardware-enforcement claims still sit at 0% reduced. They need a physical safety-module prototype to demonstrate. That's the work that opens up when we have the funding for it.

The full per-PPA breakdown is on the Investors page under the Patent Watch chart.

Why this matters

The dominant assumption in AI right now is that more compute and bigger context windows will solve memory. Stuff more tokens into the prompt; let attention figure it out. That works up to a point and then runs out. Context windows are expensive, fragile, and don't actually retain information across sessions. They just make the current conversation longer.

Real persistent memory, with real structure, with the system genuinely learning over time about what to remember and how to retrieve it, is a different shape of problem. It's the shape we've been working on. The engine is real. The integration into our own agent is in build right now. And we're open to building it into other developers' workflows when there's a fit.

Patent-pending. Christian Worldview Alignment. Built quietly in northwestern Pennsylvania.