#1 LongMemEval · beats ChatGPT 57.7% · ICLR 2025

Your AI agents keep
forgetting everything.

HippoFabric gives your agents permanent graph memory — one correction sticks forever, every session builds on the last. Works with GPT-4, Claude, Gemini, or any LLM you're already using.

#1 LongMemEval
ICLR 2025
90.6% multi-session
accuracy
10× faster than
ChatGPT
$0 API cost
self-hosted
Book a 20-minute demo See how it works →
agent_before_after.py 
# Before Luthen — resets every session
agent = YourAgent(memory=[])  # lost on restart

# After Luthen — permanent cognitive memory
from luthen import HippoFabric, AgentRunner
brain = HippoFabric(seed="your-domain")
agent = AgentRunner(brain=brain).start()
✓Memory persists across sessions · Hebbian weights active
✓Works with GPT-4, Claude, Gemini, Llama — any LLM
✓One correction → permanent behavior change · no retraining
#1LongMemEvalICLR 2025 gold standard
✓90.6%multi-session accuracy · 50%+ better than ChatGPT
✓0.46sinference speed · 10× faster
✓SOC 2 readyenterprise deployable
✓On-premisezero API cost · self-hosted

The problem every agent team hits

Every AI agent built on
RAG has the same four problems.

You've built the agent. It works in demos. But in production — every session resets, every correction evaporates, every preference disappears. That's not a memory problem. It's an architecture problem.

Session resets kill trust

Every conversation starts from scratch. Users repeat themselves. Your agent never learns who they are, what they prefer, or what they've already corrected. You're shipping a goldfish as a colleague.

session.memory = [] # every time

Vector search fails at scale

RAG finds similar text — not related concepts. "Budget" and "Q4 forecast" are deeply connected in meaning but distant in embedding space. At 100k+ documents, false positives compound silently.

cosine_similarity ≠ understanding

Corrections don't survive sessions

Your user says "never do X." Next session, it does X again. Every behavioral fix requires a retraining cycle. You're not building an intelligent agent — you're maintaining a bug list.

fine_tune(corrections, epochs=10)

Embeddings are frozen in time

Your vector store was computed once and never changes. It can't strengthen associations from use, adapt from corrections, or improve with experience. Memory that can't grow isn't memory.

embeddings.frozen = True

How HippoFabric works

A memory layer that actually thinks.

Inspired by the human hippocampus — concepts and weighted connections, activation that spreads through related ideas, memory that strengthens with use. We stopped using vectors. We built a brain.

brain.ingest()

Write to the graph

Store concepts with their relationships and weights. Every piece of knowledge joins the network and forms connections with what's already there — not embeddings, a living graph that grows.

HippoFabric · Layer 1

brain.think()

Associative recall

Activate a concept — spreading activation flows through linked ideas by weight. Not similar text. Related ideas. The right context surfaces because it's genuinely connected, not just textually close.

HippoFabric · Layer 1

brain.remember(user_id)

Permanent user memory

Every user's preferences, corrections, and history persist forever across all sessions. One call loads everything. Your agent picks up exactly where it left off — always.

Agent SDK · Layer 2

brain.correct()

Behavioral change — permanent

One correction cascades through memory, rules, and prompt templates simultaneously. No retraining. No engineers required. Permanent from the moment you call it.

Agent SDK · Layer 2

"Neurons that fire together, wire together. Luthen applies Hebbian learning to every agent interaction."

Three layers. One platform.

Each layer is useful alone.
Together, agents evolve.

Built as a stack — so you can start with just the memory layer and add the rest when you're ready.

Layer 1 · Memory

HippoFabric

For developers

The memory layer that learns from every conversation — permanently. Weighted concept edges, spreading activation, Hebbian learning. Replace your vector store.

Cognitive memory + Hebbian learning Spreading activation recall Brain seeding + sleep consolidation
Explore HippoFabric →

Layer 2 · Runtime

Agent SDK

For ML engineers

Five lines to a live cognitive agent. Memory, tools, feedback loops, cascading corrections, and sleep consolidation — wired out of the box.

AgentRunner in 5 lines Tool discovery via activation Brain cloning + fleet management
Explore Agent SDK →

Layer 3 · Governance

Cortex

For CTOs & ops

See exactly what your agents remember, decide, and do. Full audit trail, brain health monitoring, PII masking, and SafetyGate — all in a no-code console.

Business rules & scope control PII masking · SafetyGate <3ms Real-time brain health monitoring
Explore Cortex →

LongMemEval · ICLR 2025 · independent benchmark

The numbers don't lie.

Tested against ChatGPT, Claude, and Gemini on the gold standard for AI memory evaluation.

Multi-session reasoning

90.6%

50%+ better than ChatGPT
vs 57.7% · ranked #1 overall

Inference speed

0.46s

10× faster than competitors
vs 2–5s · zero API cost

Overall rank

#1

Category of one
no competitor matches memory arch.

"90.6% accuracy in multi-session reasoning — more than 50% better than ChatGPT, at 10× faster inference and zero API cost."

LongMemEval · ICLR 2025 · independent benchmark · April 2026

Emergent capabilities

Five things no other
AI memory does.

These weren't designed into the architecture. They emerged from building a memory layer that actually thinks.

01
Unique

Behavioral learning

Users teach agents conversationally. "Never use bullet points" becomes permanent behavior instantly — no retraining, no engineers, no redeployment. The correction cascades through memory, rules, and prompt templates in one call.

02

Abstract concept formation

Co-activated concepts spontaneously cluster into higher-order understanding. Your agent understands that "budget" and "Q4 forecast" are connected — without being told. It doesn't just retrieve. It reasons.

03
Unique

Transferable brains

Clone a trained brain into a new agent. It inherits every learned association, every behavioral rule, every cognitive frame — and evolves independently from there. Deploy expertise at scale.

04

Cognitive frames

Task-specific context that optimises itself through use. Prompts that stabilise into expert patterns without manual tuning. The agent gets better at the tasks it does most — automatically.

05
Unique

Sleep consolidation

Agents replay interaction traces offline, strengthen high-signal edges, and crystallise schemas — exactly like biological brains during sleep. Memory improves without any new input.

Enterprise Integration Hub

Connects to every system
you already run.

Agents speak plain language. The Integration Hub inside Cortex handles protocol, authentication, and data transformation automatically — no connector configuration, no custom code.

SalesforceSAP ServiceNowWorkday SlackJira ZendeskTeams OracleSnowflake GitHub + any system via REST
Integration Hub · live example
Agent says →
"Pull open Salesforce opportunities and cross-reference against the latest SAP forecast. Flag anything where pipeline confidence is below 60%."
Hub resolves
Salesforce — OAuth resolved · SOQL query built · 847 opportunities fetched
SAP — BAPI auth resolved · Q4 forecast extracted · cross-referenced
Result — 23 opportunities flagged · ranked by confidence · delivered to agent

Independent validation · April 2026

How does Luthen stack up?

10/10

Memory architecture
Category of one

10/10

Behavioral learning
Nobody else has this

9/10

Sleep consolidation
Unique in production

9.1

Overall score
out of 10

"Architecturally ahead of the entire field in memory & behavioral learning."

Independent validation · April 2026

Resources

AI trends, product deep-dives,
and what we're learning.

Research, guides, and thinking from the Luthen team — on cognitive AI, enterprise agents, and the future of intelligent work.

View all resources →
Research

AI Trends · Featured

Why we stopped using RAG and built a hippocampus instead

Two years building AI systems led us to one conclusion: semantic search was never designed to be a brain. Here's what we found, what we built, and why biological memory architecture changes everything for enterprise AI agents.

April 2026 12 min read Read article →
Product

Product · HippoFabric

HippoFabric vs vector stores: a benchmark breakdown

90.6% multi-session accuracy vs 57.7% for ChatGPT. How we ran the LongMemEval tests and what the numbers actually mean.

March 2026 8 min read Read article →
AI Trends

AI Trends · Enterprise

The four generations of AI agents — and why most enterprises are stuck on generation two

From rule-based bots to cognitive agents that evolve. Where the market is, where it's going, and what separates generation four from everything before it.

March 2026 10 min read Read article →
Guide

Guide · Procurement

Agentic procurement: from purchase order to managed asset

The complete guide to deploying a procurement agent network — seven stages, eight agents, one closed loop that compounds with every transaction.

April 2026 15 min read Read article →
AI Trends

AI Trends · Knowledge Work

The knowledge worker replacement thesis — what AI can actually own

70% of knowledge work is retrievable. The honest breakdown of what agents replace, what they assist with, and what genuinely requires a human.

April 2026 9 min read Read article →

Book a demo

See HippoFabric working
in your stack.

20 minutes. We'll show you persistent memory, behavioral learning, and the correction cascade — live in a real agent.

No commitment. No sales pitch. We'll show you exactly how HippoFabric solves the memory problem for your use case. Responds within 4 hours · hello@luthen.ai

© 2026 Luthen. All rights reserved.

Luthen is a registered trademark.