CORTEX is ATERNA AI's persistent memory layer for AI agents, built in-house and included with every ATERNA agent. It is modeled on human neuroscience, hippocampal memory formation, synaptic plasticity, and sleep consolidation, and gives AI agents persistent identity, compounding intelligence, and autonomous learning. It runs in production with tens of thousands of memories in a synaptic graph that grows nightly.

How is CORTEX different from RAG or vector databases?

RAG and vector databases are search systems bolted onto LLMs. CORTEX is a full cognitive architecture. Memories form synaptic connections organically. Nightly dream cycles consolidate and synthesize knowledge using a two-phase model based on human SWS and REM sleep. Agents develop real identity through accumulated experience. It is the difference between a filing cabinet and a brain.

CORTEX runs inside every agent ATERNA builds, on its own isolated instance per agent. Teams that build their own agents can license the same engine as Quartz.

MERIDIAN is ATERNA's agent framework, designed for operators, not developers. It is CORTEX-native from day one, meaning every agent has full persistent memory, dream cycles, and identity emergence built in.

What is the ATERNA AI Workforce?

ATERNA operates a 9-agent autonomous AI workforce, all powered by CORTEX. Each agent has a specialized role, strategy, engineering, sales, content, research, operations, client success, and persistent memory. They coordinate, learn, and compound intelligence over time. This workforce is the living proof that CORTEX works in production.

Cognitive memory, measured.

Name: CORTEX
Author: ATERNA AI

Reproducible scores across the benchmarks that actually measure agent memory. Open methodology. Pick your own benchmark. Submit your own score.

8 systems, 6 benchmarks. CORTEX V2.4, run 2026-04-22T22:22Z.

Fig. 1 The leaderboard, sorted by system name ↑Click a column head to sort. The leader in each column is set in amber.


Atlan MCP metadata, text-to-SQL	–⁶	–⁶	–⁶	–⁶	–⁶	–⁶
Cognee	–	–	–	–	–	–
CORTEX V2.4, ATERNA	100.0% reproduce	93.6% reproduce	78.5% reproduce	pending⁷	40.1%⁸ reproduce	pending⁹
HippoRAG 2 OSU + UIUC	–	–	–	–	–	–
Letta formerly MemGPT	–	74.0% source	–	–	–	–
Mastra Observational Memory	94.9%¹ source	–	–	–	–	–
Mem0	93.4%¹ source	91.6%¹ source	–	no lift² source	64.1% source	decline³ source
Zep Graphiti	+18.5%⁴ source	58.4%⁵ source	–	no lift² source	–	–

Sort by

Atlan MCP

metadata, text-to-SQL

LongMemEval oracle, R@10, saturated: –⁶
LoCoMo retrieval, R@10: –⁶
CogBench composite: –⁶
MEMTRACK multi-source state: –⁶
BEAM 1M token: –⁶
HaluMem op-level hallucination: –⁶

Cognee

LongMemEval oracle, R@10, saturated: –
LoCoMo retrieval, R@10: –
CogBench composite: –
MEMTRACK multi-source state: –
BEAM 1M token: –
HaluMem op-level hallucination: –

CORTEX

V2.4, ATERNA

LongMemEval oracle, R@10, saturated: 100.0%
reproduce
LoCoMo retrieval, R@10: 93.6%
reproduce
CogBench composite: 78.5%
reproduce
MEMTRACK multi-source state: pending⁷
BEAM 1M token: 40.1%⁸
reproduce
HaluMem op-level hallucination: pending⁹

HippoRAG 2

OSU + UIUC

LongMemEval oracle, R@10, saturated: –
LoCoMo retrieval, R@10: –
CogBench composite: –
MEMTRACK multi-source state: –
BEAM 1M token: –
HaluMem op-level hallucination: –

Letta

formerly MemGPT

LongMemEval oracle, R@10, saturated: –
LoCoMo retrieval, R@10: 74.0%
source
CogBench composite: –
MEMTRACK multi-source state: –
BEAM 1M token: –
HaluMem op-level hallucination: –

Mastra

Observational Memory

LongMemEval oracle, R@10, saturated: 94.9%¹
source
LoCoMo retrieval, R@10: –
CogBench composite: –
MEMTRACK multi-source state: –
BEAM 1M token: –
HaluMem op-level hallucination: –

Mem0

LongMemEval oracle, R@10, saturated: 93.4%¹
source
LoCoMo retrieval, R@10: 91.6%¹
source
CogBench composite: –
MEMTRACK multi-source state: no lift²
source
BEAM 1M token: 64.1%
source
HaluMem op-level hallucination: decline³
source

Zep

Graphiti

LongMemEval oracle, R@10, saturated: +18.5%⁴
source
LoCoMo retrieval, R@10: 58.4%⁵
source
CogBench composite: –
MEMTRACK multi-source state: no lift²
source
BEAM 1M token: –
HaluMem op-level hallucination: –

1.LLM-judge accuracy metric. Not directly comparable to retrieval R@10. See the note on metrics below.
2.Per the MEMTRACK paper (Patronus AI, NeurIPS 2025): Zep and Mem0 do not significantly improve LLM performance on this benchmark. arXiv:2510.01353
3.Per the HaluMem paper (MemTensor, Nov 2025): Mem0, Mem0-Graph and Memobase all show marked decline on long-context variants. arXiv:2511.03506
4.Zep reports percentage lift over baseline only; absolute R@10 unpublished. arXiv:2501.13956
5.Zep's original 84% LoCoMo claim corrects to 58.44% per issue #5 on their public zep-papers repo. We list the corrected number.
6.Atlan's published evaluations focus on metadata-grounded text-to-SQL accuracy (reported 3x improvement over a non-Atlan baseline). No agent-memory benchmark scores published.
7.CORTEX on MEMTRACK: the adapter is complete. The score is withheld until Patronus releases the official dataset, so the comparison against Mem0 and Zep runs on the same data.
8.CORTEX on BEAM: the official BEAM-1M set (35 conversations, 700 questions). Retrieval, then Haiku 4.5 synthesis, then a Haiku 4.5 judge. The raw judged run is linked from the cell.
9.CORTEX on HaluMem: the adapter is planned and no score exists yet. The cell fills when a run has been published under /bench.

What each benchmark measures.

Six benchmarks, six different questions. Each one below names its paper, what it tests, and why it is on the board.

LongMemEval oracle, R@10, saturated

Wu et al., ICLR 2025

500 long-conversation questions across 5 question types. Tests retrieval plus single-hop reasoning. Saturated for retrieval systems in 2026 but still the canonical entry point.

ii.

LoCoMo retrieval, R@10

Maharana et al., ACL 2024

10 multi-session dialogues averaging 300 turns. Categories: single-hop, multi-hop, temporal, open-domain. Published retrieval scores are comparable across systems.

iii.

CogBench composite

ATERNA, 2026

Cognitive-memory capability suite: temporal validity, reconsolidation, novelty detection, emotional recall, cross-agent transfer, compounding intelligence, procedural learning. Designed for memory systems, not LLM context.

iv.

MEMTRACK multi-source state

Patronus AI, NeurIPS 2025

Slack, Linear and Git event streams. Correctness, efficiency, redundancy. Zep and Mem0 do not significantly improve GPT-5 on this benchmark, per the paper.

BEAM 1M token

Tavakoli et al., ICLR 2026

First benchmark with single-user narrative coherence at 1M and 10M tokens. Ten abilities including contradiction resolution and event ordering. The LIGHT framework uses an explicit working-memory component.

vi.

HaluMem op-level hallucination

MemTensor, Nov 2025

First operation-level hallucination benchmark for memory systems. Three stages: extraction, updating, QA. Mem0, Mem0-Graph and Memobase all show marked decline on long context, per the paper.

On metrics: LLM-judge scores (Mem0, Mastra, Supermemory on LongMemEval) and retrieval R@10 scores (CORTEX) are not directly comparable. Retrieval recall asks "did the system surface the right memory." LLM-judge asks "did the end-to-end answer match ground truth." CORTEX prioritizes retrieval R@10 because a memory system that cannot surface the right memory cannot be made correct by any downstream LLM. We publish both when available.
On saturation: LongMemEval oracle top-10 is saturated for 2026-era retrieval systems. That is why we are shipping BEAM, MEMTRACK and HaluMem adapters: discriminating benchmarks that test state tracking, cross-source coherence and operation-level hallucination rather than session-level retrieval accuracy.
On competitor data: We do not re-run competitor systems. Where a competitor has published a score, we cite it with a source link. Where they have not, the cell shows a dash. Corrected numbers (Zep LoCoMo) are shown with a footnote and a link to the correction thread.

Run a memory system we missed? Post a score.

Send raw JSON and reproduction instructions to benchmarks@aterna.ai. We list every verifiable result, including ones that beat CORTEX. This is a public scoreboard, not a marketing page.

Benchmark snapshot

github.com/ATERNA-AI/cortex

Archived at v2.4 so these scores stay reproducible. Current CORTEX ships inside ATERNA agents.

Reproduce locally

git clone https://github.com/ATERNA-AI/cortex
cd cortex && npm install
npx tsx benchmarks/longmemeval/run.ts --topk 10 --dataset oracle
npx tsx benchmarks/locomo/run-retrieval.ts --topk 10
npm run cogbench:run

Dataset downloads and the full steps are in BENCHMARKS.md.

Build

CORTEX V2.4

run 2026-04-22T22:22Z

raw longmemeval.json, locomo.json, cogbench.json

See the memory behind the numbers.

In a 20-minute live demo we show CORTEX recalling real client context, then map what it would do for your team.

Book a live demo