For developers building AI agents

Memory for AI agents that understands — not just stores.

Cognitive Primitives System is a structural memory-and-learning engine for AI agents. It remembers by meaning, not text similarity — and drops into LangChain, CrewAI, or AutoGen in a few lines.

Apply for Early Access

Working system today — private beta open, live demo coming.

The problem

If you've shipped an agent, you've hit this.

Your agent nails a task today and remembers nothing useful tomorrow. A fix it found last week never carries over to a structurally identical problem this week. Pile on more history and retrieval gets worse, not better.

The longer your agent runs, the worse it gets.

How it works

Three calls. Any LLM. No retraining.

Cognitive Primitives System drops into your existing memory slot — no GPUs, no fine-tuning, LLM-agnostic.

from cognitive_primitives import Memory

mem = Memory()                       # drop-in — works with your LLM

# SAVE — turn a raw agent experience into a structured concept card
mem.structurize(agent_experience)

# FIND — retrieve by structure, not keyword similarity
hits = mem.query("enterprise client threatening to churn")

# CONNECT — ask how two cases relate, structurally
mem.relate(case_a, case_b)

Illustrative — the final API may change before public release.

structurize() — save any experience as a structured concept card, automatically.
query() — retrieve the structurally relevant past experience, ranked.
relate() — ask how any two cases or ideas connect, in structural terms.

3–5 lines in your existing stack. It compounds with use — the longer it runs, the smarter your agent gets.

Why it's different

Not another vector store.

Mem0, Letta, and LangMem save text and search it by similarity — a pile of sticky notes with a good search bar. That finds similar-sounding memories, not the right ones.

A vector search reads “billing error, customer furious” and “damaged product, customer wants a refund” as unrelated. Structurally, they're the same case: control was taken away, emotion is high, and the fix is to restore control. Cognitive Primitives System matches on that structure — so your agent recalls what actually worked, not what merely sounds alike.

We're not a better vector store. We're the layer above it.

Early Access · First Cohort

Build this with us.

We're onboarding a small first cohort of agent builders who've hit the ceiling of vector memory. This isn't a waitlist for a finished product — it's an invitation to shape one.

Free access throughout the private beta.
A direct line to the founder — your feedback goes straight into the roadmap.
Founding-member pricing, locked in for life when paid tiers launch.
First to ship structural memory in your agent — before anyone else has it.

You're a fit if you're building agents on LangChain, CrewAI, AutoGen, or LlamaIndex — and tired of your agent forgetting what matters and repeating the same mistakes.

FAQ

Questions before you apply.

Which frameworks does it support?

LangChain, CrewAI, AutoGen, and LlamaIndex at launch — with a framework-agnostic SDK underneath, so you can wire it into anything.

Which LLMs does it work with?

Any of them. We're LLM-agnostic — bring your own. The LLM is only used at the boundary; the core runs without it.

Is my data private?

Your concept cards are yours. We don't train on your data. (Full privacy policy to be published before public launch.)

What does early access cost?

Nothing during the private beta. Founding users keep founding-member pricing when paid tiers launch.

When's the live demo?

Weeks away — early-access users get to see it first.