The most accurate memory for AI agents. Secure, governed and built to scale.
CortexDB gives AI products and agents a long-term memory of their users, customers and work. It holds the highest published LongMemEval score of any production memory system, and it was built for the enterprise from day one: permissions checked on every memory it returns, erasure you can verify, a replicated cluster, and one binary from a laptop to production.
- Run blindNo benchmark labels or answers ever reach the server.
- Released buildThe 0.14.2 image customers and CortexDB Cloud run.
- Official judgeLongMemEval’s own evaluator, not a friendlier one.
- ReproducibleConfig, harness and outputs ready for third-party reruns.
Number one on the standard memory benchmark
LongMemEval asks 500 questions about long chat histories: facts the user stated, what changed, what happened when, what the assistant said, and what the user prefers. CortexDB 0.14.2 answers 95.4% correctly (477 of 500) under the benchmark’s own judge, higher than any production memory system has published. It was run blind, on the released build, with a configuration we hand to anyone who wants to check it.
| System | Accuracy | Judged by |
|---|---|---|
| CortexDB 0.14.2Claude Opus 5.5 reader | official judge · blind | |
| Mem0 Platformreader not disclosed | judge not disclosed | |
| EverOS (EverMind)method not stated | method not stated | |
| Mastra Observational Memorygpt-5-mini reader | official judge | |
| Backboardreader not stated | best of 4 judges | |
| MemMachinegpt-5-mini reader | gpt-4o-mini judge | |
| HonchoGemini 3 Pro reader | judge not named | |
| Lumetra EngramGPT-5 reader | gpt-4o judge | |
| Hindsight (Vectorize)Gemini 3 Pro reader | open-model judge | |
| Zepgpt-5.4 reader | gpt-5.4 judge | |
| SupermemoryGemini 3 Pro reader | official judge | |
| Full chat history in GPT-4ono memory system | official judge |
Competitor figures are their own published results (sources below). Readers (the model that writes the answer) and judges differ between vendors, as marked. Mastra’s headline 94.87% averages its category scores; 93.6% is its overall accuracy from its published counts. Experimental and retrieval-only claims are not charted; the sources list them.
CortexDB 0.14.2 by question type
| Question type | Questions | Correct |
|---|---|---|
| Facts the user stated | 70 | 100% |
| What the assistant said | 56 | 100% |
| Knowledge that changed | 78 | 100% |
| User preferences | 30 | 93.3% |
| Across many sessions | 133 | 93.2% |
| Reasoning about time | 133 | 91.0% |
| All questions | 500 | 95.4% |
Why this number holds up
- Run blind. The harness strips question types, IDs and answers before anything reaches the server.
- The benchmark’s own judge. LongMemEval’s evaluator with GPT-4o, not a custom or friendlier grader.
- Released software. The 0.14.2 image customers run, with a configuration we hand to any third party.
- Repeatable. The previous release, 0.14.1, scored 95.8% under the same judge a day earlier.
Long conversations: LoCoMo
LoCoMo asks questions about ten very long conversations, each about 600 turns over up to 32 sessions. On all 1,540 questions in categories 1–4, CortexDB 0.14.2 answers 92.8% correctly (1,429) under a binary LLM judge. It ran label-blind: every question gets the same prompt, and no category reaches the server. Category 5, adversarial questions, is not scored, as in Mem0’s protocol.
LoCoMo categories 1–4 (1,540 questions). CortexDB 0.14.2 answering with Claude Opus 5.5, scored by a gpt-4.1 binary LLM judge using the Memori/Mem0 prompt, Nov 5, 2026.
Fast to write, light to run
The same LongMemEval run, instrumented end to end: 466,283 memories (1.02 million indexed chunks) written into 500 separate memory spaces and 500 questions answered, on one 16-core server. Every number below is from that run.
median to commit a batch of ~880 memories to the replicated log (p99 19 ms)
of RAM per memory: 5.3 GB peak for all 466K memories
per memory, everything included: log, vectors, text index, layers
average over the whole run (1.4 peak): the work waits on the model APIs, not on CortexDB
Latency (milliseconds)
| Stage | p50 | p95 | p99 |
|---|---|---|---|
| Durable write (batch of ~880) | 12 | 15 | 19 |
| Admission | 7 | 10 | 16 |
| Searchable, incl. remote embedding | 12,142 | 18,550 | 23,141 |
| Recall for an answer | 2,356 | 8,830 | 12,752 |
| Answer generation (Opus 5.5) | 2,983 | 15,262 | 21,037 |
| Full answer, end to end | 5,615 | 23,246 | 29,073 |
A write is durable as soon as the log commits it. It becomes searchable once its embeddings come back from the embedding provider, which is where nearly all of that time goes. Recall here includes the parallel sub-queries and reranking an answer runs.
Where the 8.9 GB on disk goes
- Vector index 6.74 GB
- Replicated and derived logs 0.98 GB
- Text index 0.33 GB
- Content, history and layers 0.84 GB
What a question costs
$0.095 in answer-model tokens per question with Claude Opus 5.5 at high effort, every call included (21.8K input tokens on average), and $1.59 to embed all 80 million tokens of the 500 histories. You choose the models; CortexDB’s usage ledger records the tokens and cost of every call.
CortexDB 0.14.2 release image on one AWS m7i.4xlarge (16 vCPU, 64 GB). Embeddings: OpenAI text-embedding-3-small. Reranking: Cohere rerank-v3.5. Eight questions in flight at once. RAM is the server process’s resident memory; disk is the whole data directory after the run.
How CortexDB compares
Memory layers, enterprise context platforms, a hyperscaler memory service and an authorization engine, side by side on what an enterprise asks. CortexDB is the only product with a ✓ in every row.
| Feature | CortexDB | Mem0 | Zep | Letta | Supermemory | Hindsight | Glean | Scuba | Engram | AWS AgentCore | AuthZed |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Memory quality | |||||||||||
| LongMemEval-S, published (%) | 95.4 | 94.8* | 90.2* | not documented | 85.2 | 91.4* | not documented | not documented | not documented | not documented | not applicable |
| Facts over time, contradictions handled | documented | partial, preview, plan-gated or marketing only | documented | partial, preview, plan-gated or marketing only | documented | documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | not documented | partial, preview, plan-gated or marketing only | not applicable |
| Keeps the raw history, not only extracts | documented | partial, preview, plan-gated or marketing only | documented | documented | documented | documented | partial, preview, plan-gated or marketing only | documented | partial, preview, plan-gated or marketing only | documented | not applicable |
| Multilingual memory | documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | not documented | partial, preview, plan-gated or marketing only | documented | documented | not documented | not documented | documented | not applicable |
| Scale and deployment | |||||||||||
| On-prem or air-gapped | documented | documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | documented | documented | not documented | not supported | not documented | not supported | documented |
| Replicated cluster (high availability) | documented | not documented | partial, preview, plan-gated or marketing only | not documented | not documented | documented | documented | not documented | not documented | partial, preview, plan-gated or marketing only | documented |
| Sharding with online tenant moves | documented | not documented | partial, preview, plan-gated or marketing only | not documented | not documented | documented | not documented | not documented | not documented | not applicable | not documented |
| Old memory tiered to object storage | documented | not documented | partial, preview, plan-gated or marketing only | not documented | not documented | documented | not documented | not documented | not documented | not documented | not applicable |
| Organization and access | |||||||||||
| Org hierarchy, shared and private memory | documented | partial, preview, plan-gated or marketing only | documented | documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | not documented | partial, preview, plan-gated or marketing only | not applicable |
| Fine-grained permissions (ReBAC) | documented | not documented | documented | partial, preview, plan-gated or marketing only | documented | partial, preview, plan-gated or marketing only | documented | documented | not documented | documented | documented |
| SCIM user provisioning | documented | not documented | partial, preview, plan-gated or marketing only | not documented | not documented | not documented | not documented | not documented | not documented | partial, preview, plan-gated or marketing only | not documented |
| Scoped tokens, delegation to agents | documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | documented | partial, preview, plan-gated or marketing only | documented | partial, preview, plan-gated or marketing only | not documented | partial, preview, plan-gated or marketing only | not applicable |
| AuthZEN 1.0 authorization API | documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented |
| MCP server with authorization | documented | documented | documented | partial, preview, plan-gated or marketing only | documented | documented | documented | documented | not documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only |
| Governance | |||||||||||
| Right-to-erasure API | documented | documented | documented | documented | documented | documented | partial, preview, plan-gated or marketing only | partial, preview, plan-gated or marketing only | not documented | partial, preview, plan-gated or marketing only | not applicable |
| Verifiable erasure (receipt and proof) | documented | not documented | not documented | not documented | not documented | documented | not documented | not documented | not documented | not documented | not applicable |
| Legal hold | documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented | not documented | not applicable |
| Retention policies | documented | partial, preview, plan-gated or marketing only | documented | not documented | partial, preview, plan-gated or marketing only | documented | documented | partial, preview, plan-gated or marketing only | not documented | documented | not applicable |
| Audit log | documented | partial, preview, plan-gated or marketing only | documented | not documented | partial, preview, plan-gated or marketing only | documented | documented | not documented | not documented | documented | not documented |
- documented
- partial, preview, plan-gated or marketing only
- not supported
- not documented
- not applicable
- judge not disclosed or not the benchmark’s own
Competitor cells summarize public documentation retrieved November 5, 2026, from a 23-product study in which every cell carries a source; “not documented” is not proof a feature is missing. A row that joins two features shows ◐ when only one is documented. CortexDB’s column describes release 0.14.2.
AuthZed (SpiceDB) is an authorization engine: it decides who may see what but stores no memories, so memory rows do not apply; its MCP servers are developer tools. Engram is engram.com, a pre-release lab, not Lumetra Engram. Zep’s 90.2% uses a GPT-5.4 judge; its paper reported 71.2% under the official judge. Hindsight’s 91.4% is from its paper.
Why teams choose CortexDB
Accuracy you can check
95.4% on LongMemEval, run blind on the released build under the benchmark’s own judge. We give you the image, config and harness to rerun it.
Every memory permission-checked
Scoped tokens, org trees with private regions, your identity provider’s tokens, SCIM 2.0 and AuthZEN 1.0. Each memory is checked before it reaches the model.
Erasure you can prove
One request erases a person or tenant across the log, indexes, replicas and object tier, then verifies it with a receipt. Legal holds and retention built in.
One system, not three
No separate authorization engine and sync jobs: memories, permissions and deletions live in one replicated log.
Built to scale
Raft-replicated shards with 1 to 5 replicas, online tenant moves, and old memories tiered to object storage, still searchable.
Runs where you need it
CortexDB Cloud, your cloud or data center, air-gapped. Python and TypeScript SDKs, MCP, 18 connectors, bring your own models.
How we measured, and where every number comes from
CortexDB LongMemEval-S. CortexDB 0.14.2 release image on AWS m7i.4xlarge; blind harness from the v0.14.2 tag; LongMemEval-S cleaned (500 questions); reader Claude Opus 5.5 (adaptive thinking, high effort); recall budget 8,000 tokens; Cohere rerank-v3.5; OpenAI text-embedding-3-small. The harness also extracts facts from every session with gpt-4o-mini on the client and stores them beside the raw turns, the same for every question. Scored with LongMemEval’s evaluate_qa.py and GPT-4o (gpt-4o-2024-08-06): 477/500. One question returned no answer and is counted wrong. 0.14.1, same setup, Nov 4: 479/500. cortexdb.ai/docs/operations/benchmarking
Mem0. 94.8% (top-50) / 94.4% (top-200), Platform, Apr 2026, reader and judge not disclosed. mem0.ai/blog/the-token-efficient-memory-algorithm-now-has-temporal-reasoning
EverOS. 94.00%, homepage, method not stated. evermind.ai
Mastra Observational Memory. 94.87% average of six categories; 93.6% overall (468/500) from published counts; gpt-5-mini reader, GPT-4o judge, Feb 2026. mastra.ai/research/observational-memory
Backboard. 93.4% (467/500), highest of four unnamed judges, Feb 2026. backboard.io/blog
MemMachine. 93.0%, GPT-4o-mini judge, GPT-5-mini reader, arXiv Apr 2026. arxiv.org/abs/2604.04853
Honcho. 92.6%, Gemini 3 Pro, judge not named, Dec 2025. honcho.dev/evals
Lumetra Engram. 91.6% (458/500), GPT-4o grade, Apr–May 2026. lumetra.io/engram-on-longmemeval
Hindsight. 91.4%, GPT-OSS-120B judge, Gemini-3 Pro reader, arXiv Dec 2025. arxiv.org/abs/2512.12818
Zep. 90.2% (451/500), GPT-5.4 judge and reader, Sep 2026; 71.2% under the official judge with GPT-4o, arXiv Jan 2025. getzep.com/research arxiv.org/abs/2501.13956
Supermemory. 85.2%, GPT-4o judge, Gemini 3 Pro reader. supermemory.ai/research/longmembench
Full history in GPT-4o. 60.6%, LongMemEval paper (ICLR 2025). arxiv.org/abs/2410.10813
Not charted. Retrieval-recall results, which do not measure answers (MemPalace 96.6% R@5, Supermemory’s Recall@20 table); Supermemory ASMR, experimental and not its production engine, no code published (98.6% if any of 8 answers is right, 97.2% from a 12-agent ensemble); Hindsight’s site figure of 94.6% “retrieval accuracy”, undefined.
CortexDB LoCoMo. CortexDB 0.14.2 answering on the server through /v1/answer, with Claude Opus 5.5 and a 7,000-token recall budget; all 10 conversations, categories 1–4 (1,540 questions); label-blind, with the same prompt for every question and no category sent to the server. Scored by a binary CORRECT/WRONG LLM judge, gpt-4.1, with the Memori/Mem0-style prompt (not F1 or BLEU): 1,429/1,540. Category 5 (adversarial) is not scored, as in Mem0’s protocol. Nov 5, 2026. cortexdb.ai/docs/operations/benchmarking
LoCoMo. Maharana et al., “Evaluating Very Long-Term Conversational Memory of LLM Agents”, Snap Research; code and data on the project page. snap-research.github.io/locomo
Feature comparison. 23 products, 54 sub-dimensions, 1,242 cells, each with a source, from vendor documentation, pricing, security pages and repositories retrieved November 5, 2026. Absence of documentation is shown as “not documented”, not as missing. AuthZed’s column comes from a separate study of nine authorization engines. CortexDB’s column describes release 0.14.2 from its own documentation.
Performance. Server request log (every request), the indexer’s phase timings, cgroup and process memory sampled every 10 s, and a per-store disk census every 60 s, all on the LongMemEval run above. Model spend is from the server’s own usage ledger, which records every call.
Check the number on your own hardware.
We hand over the exact image digest, server configuration and blind harness behind the 95.4%. Run them with your own API keys and score the answers with LongMemEval’s official evaluator. Or start free on CortexDB Cloud and try it on your own data.