CortexDB V1: The Experience Layer for AI Agents

The missing third layer

Today's AI agents are amnesiacs. Open ChatGPT, tell it your name, log out, come back tomorrow — it has no idea who you are. For a chatbot that's annoying. For an agent — one that books travel, manages a customer pipeline, debugs production systems — it's broken.

People have tried to paper over this by stuffing more text into the prompt. The prompt is a clipboard, not a memory. You can't enforce "delete this customer's data" on a clipboard. You can't audit who saw what. You can't answer why do you think the deal is at risk? from a clipboard.

A real AI agent needs three layers, stacked together:

Layers 1 and 2 are commodities. Layer 3 is the moat — and Layer 3 is what CortexDB builds.

Memory isn't storage. It's a cycle.

Real memory — the kind your own brain runs — isn't a single operation. It's a continuous cycle of operations the brain performs, over and over. Skip any one and you're back to a vector DB.

Capture  ──▶  Extract  ──▶  Reconcile  ──▶  Forget  ──▶  Consolidate
 (WAL)       (facts +       (bi-temporal     (lifecycle    ("sleep" —
              entities)      supersession)    methylation)   the brain)

1. Capture. Ingest raw experience. Conversations, events, observations — preserved exactly as they happened in a tamper-proof WAL.
2. Extract. Pull facts, entities, relationships out of raw events. Turn experience into structured knowledge.
3. Reconcile. When memories contradict — Bob said yes Tuesday, no Friday — decide what's true now. Newest wins, with the older version preserved as true until Tuesday.
4. Forget. Prune the irrelevant. Not all memories matter. The brain forgets on purpose.
5. Consolidate. When the agent "sleeps," build a coherent worldview from raw events. Synthesis. Generalization. Procedure formation. This is where memory becomes intelligent.

V1 ships the first four stages as stable surface. Consolidation is in flight as a beta endpoint — POST /v1/understanding/synthesize.

The five layers

If the cycle is what memory does, the layers are what memory contains. Five layers, stacked from raw to refined, each addressable through its own API endpoint:

A system with only Events is a log. A system with only Facts is a database. A system with all five layers, with the right relationships between them, is the experience layer that completes the agent.

Almost no AI memory product handles Beliefs properly. They have facts. They don't track confidence or evidence. They can't answer "why do you think that?" — and that question is one of the biggest reasons enterprise buyers don't trust agents yet. CortexDB can: GET /v1/beliefs/why?belief_id=… returns the full support graph plus a narrative rendering.

The benchmarks

Anyone can claim great AI memory. There are exactly two public, standardized benchmarks the research community has rallied around — LongMemEval-S and LoCoMo — and V1 hits state-of-the-art on both.

The combined trajectory: +17.4 points on LongMemEval-S and +10.6 points on LoCoMo in six weeks of focused engineering. Velocity is the second-order signal of architecture quality — most teams in this space spent six weeks integrating yet another reranker. We've been rebuilding the foundation.

Why nobody else is doing this

The market has been arguing about which existing piece is the answer to AI memory. None of them is — alone or together.

Stitching these together with prompts and orchestration code is what every team is doing today — and it's why no AI yet feels like it actually knows you. Trying to build a brain by gluing a search engine to a Rolodex to a filing cabinet to a phone book.

The peer attempts make the same architectural mistake — they build memory as a feature on top of a vector store. Mem0 stores LLM-rewritten snippets and ranks them at recall time; they can't tell you why a result ranked where it did because the derivation isn't stored. Zep separates facts from episodes but folds Belief into Fact, losing the distinction between what I observed and what I concluded. Letta has tiered memory but no explicit Belief layer at all. We've broken those layers apart on purpose — because the trail back to evidence is the entire point.

What V1 ships

The full V1 surface is about ~30 stable endpoints plus an SSE channel. The headline pieces:

Eight weeks from first commit

First commit landed on March 16, 2026. V1 ships today, May 16. In eight weeks:

Most startups in this space are pre-product or pre-benchmark. We're eight weeks in and have already won the public benchmarks twice. Architecture is the multiplier. When the foundation is right, adding a feature isn't a project — it's a commit.

What's next: V2 — the brain

V1 ships the four stages of the memory cycle that operate continuously: Capture, Extract, Reconcile, Forget. V2 is about the fifth — Consolidate. The "sleep" stage where the brain shows up. Where raw experience becomes understanding.

V2's targets: 96–97% on LongMemEval-S, 92–94% on LoCoMo, and p50 end-to-end answer latency under 4 seconds. The synthesizer is already in flight as a beta endpoint — call POST /v1/understanding/synthesize to trigger a consolidation pass over a scope today, and the concepts land in GET /v1/understanding.

Even the primitive version of the "sleep" stage is unique in the market. No competitor builds it. That's what makes CortexDB feel like a brain instead of a log — and the gap widens with every cycle.

curl -X POST https://api-v1.cortexdb.ai/v1/auth/signup -d '{}'

The bet

AI is having its internet moment. Reasoning got cheap. Knowledge retrieval is becoming infrastructure. What's left — and what every operator and every end-user actually wants — is an AI that knows them, knows their business, learns from every interaction, and gets sharper because of it.

That AI doesn't exist yet. To exist, it needs an experience layer underneath: durable, queryable, governed, explainable. The third stack that's been missing.

We're building it. The first version ships today. The next version ships in weeks. We'd love to have you on for the ride.