CogmemAi Scores 95.1% on LongMemEval: The Field’s Top Published Score
The Result
On April 19, 2026, CogmemAi scored 95.1% on LongMemEval — 97 out of 102 questions correct. It is the top published score on the benchmark.
The paper that introduced LongMemEval reported its best baseline at 64%. Most memory systems in production today score between 50% and 70%.
CogmemAi is the Smart Persistent Memory layer for AI agents. It is available today as an MCP server, a REST API, and a direct API — free tier included.
What Is LongMemEval?
LongMemEval is the benchmark researchers reach for when they want to know whether an AI system can actually remember. It was designed to stress-test long-term memory in conversational AI over hundreds of turns spanning weeks of interaction — not the trivial 2-turn recall that most "memory" demos show.
The benchmark measures five capabilities:
- Single-session recall — remembering something said in one conversation
- Multi-session recall — remembering something said across many conversations
- Temporal reasoning — answering questions that require understanding when things happened
- Knowledge updates — updating memory when the user says something has changed
- Abstention — knowing when not to answer because the information was never in memory
These are the exact failure modes that make most AI agents feel like goldfish. LongMemEval catches them all.
How The Scores Were Verified
We publish our methodology so the numbers are verifiable:
- LoCoMo — 91%, scored April 2, 2026. 100 questions from the LoCoMo dataset, 100% retrieval hit rate, above the 87.9% human baseline on the benchmark.
- LongMemEval — 95.1%, scored April 19, 2026. 102 questions (LME-100 subset plus two additional), 97 correct, 5 incorrect.
- Answering model: Claude Opus 4.7
- Judge model: GPT-4o
The two scores are 17 days apart. In those 17 days several improvements shipped — so the LoCoMo number reflects the system as it was on April 2, and the LongMemEval number reflects the system as it was on April 19. We have not re-run LoCoMo on the current build. If we did, we expect the number would move up.
We do not share the internal configuration of CogmemAi itself — our retrieval engine, storage strategy, and reasoning layer are the work product, and we keep them private. The scores and the methodology are public. The implementation is ours.
Why This Matters For AI Agents
Every AI agent you build today loses memory at the session boundary. Your coding assistant forgets your architecture the moment the context window fills. Your customer-support agent forgets a caller between tickets. Your autonomous system forgets everything the operator rotated out with.
Memory is the primitive that separates an AI that impresses you in a demo from an AI that earns a seat on your team. A 95.1% Smart Persistent Memory system means:
- A coding agent that remembers last week’s decisions without being re-told
- An analyst agent that retains a pattern observed across thirty sessions
- An autonomous platform that inherits its predecessor’s observations
- A personal assistant that knows your preferences without being asked again
At 64% — the paper’s best baseline — the agent is wrong often enough that you can’t trust it. At 95.1%, the agent is reliable enough to be worth building on.
Who Is CogmemAi For?
CogmemAi is used by three kinds of teams:
Developers building AI agents
Drop-in Smart Persistent Memory for Claude Code, Cursor, Windsurf, Cline, Continue, and any MCP-compatible tool. Your agent remembers your codebase, your preferences, and your past decisions across every session. Free tier, 60-second install.
Defense and autonomy platforms
Memory inheritance across assets, crews, and shifts. When operators rotate, vehicles rotate, or sensors change, the system’s understanding compounds instead of resetting. Dedicated on-prem deployment available — your Smart Persistent Memory infrastructure runs entirely inside your environment, with no data crossing your network boundary. US citizen and clearance-ready engineering.
Enterprise agent teams
Customer support, sales, legal, clinical, and analyst workflows all depend on long-horizon memory. CogmemAi slots under your agent layer without replacing it — it is infrastructure, not a platform.
How To Add CogmemAi To Your Agent
Four integration paths:
MCP server (Claude Code, Cursor, Windsurf, Cline, Continue)
npm install -g cogmemai-mcp
npx cogmemai-mcp setupThe setup wizard writes your MCP config, registers your free API key, and you are done. Sixty seconds from install to first memory saved.
HTTPS REST API
Language-agnostic. Every memory operation — save, recall, update, link, consolidate — is a single HTTP call. Works with any stack.
Direct API
For teams integrating CogmemAi deep into an existing autonomy or agent stack. Lower latency, tight coupling, built-in for MCP-native and non-MCP workflows alike.
On-Prem Dedicated Deployment
For defense, government, and regulated customers, CogmemAi ships as a dedicated deployment inside your environment — no shared infrastructure, no external data egress, no third-party dependencies on your network. Your Smart Persistent Memory stays on your network boundary.
Frequently Asked Questions
What is the best AI memory benchmark score?
CogmemAi holds the top published score on LongMemEval at 95.1%, measured April 19, 2026. On LoCoMo, CogmemAi scored 91% on April 2, 2026 — above the 87.9% human baseline.
Why is the LongMemEval score higher than the LoCoMo score, if LoCoMo is the easier benchmark?
The LoCoMo run happened on April 2, 2026. The LongMemEval run happened on April 19, 2026. Between those two dates, several improvements shipped. The LoCoMo number reflects the system as it was on April 2 — the current system would likely score higher. We will re-run LoCoMo when we have a reason to. The LongMemEval result is the one that matters more for long-horizon agent work.
How do I add memory to Claude Code?
Install the CogmemAi MCP server: npm install -g cogmemai-mcp then run npx cogmemai-mcp setup. The wizard registers your MCP config automatically. Free tier available.
How do I add memory to Cursor or Windsurf?
Both support MCP servers. Install CogmemAi the same way — npm install -g cogmemai-mcp — then point your Cursor or Windsurf MCP config at it. Per-editor guides at hifriendbot.com/developer.
Is CogmemAi free?
Yes — the free tier includes 500 memories and 500 extractions per month across 5 projects. No credit card required.
Is CogmemAi open source?
The MCP client is available on npm. The core Smart Persistent Memory infrastructure — retrieval engine, storage strategy, reasoning layer — is proprietary. This is how we can guarantee benchmark performance and keep improving.
How does CogmemAi compare to Letta, mem0, Zep, Memgpt?
Benchmark scores across vendors are difficult to compare because every vendor runs their own evaluation with different models and configurations. What we can say: we published our methodology, our answering model, our judge model, and our scores. We did this on purpose — so you can trust the numbers.
Does CogmemAi offer on-prem deployment?
Yes. For defense, government, and regulated customers, CogmemAi ships as a dedicated on-prem deployment — your Smart Persistent Memory infrastructure runs entirely inside your environment, with no data crossing your network boundary. Contact us through the developer page to discuss a pilot.
Does CogmemAi work offline?
CogmemAi offers cloud, local, and hybrid storage modes. Cloud uses the full intelligence engine and scores highest on retrieval quality. Local keeps your data on your machine. Hybrid combines both with automatic fallback.
Is CogmemAi secure?
Yes. CogmemAi uses quantum-safe encryption on stored memories and supports scoped API keys, team isolation, and per-project access control. Full technical details at hifriendbot.com/tech.
Try It
If your agent forgets everything between sessions, give CogmemAi a try.
Get your free API key — no credit card, 60-second setup, works with Claude Code, Cursor, Windsurf, Cline, and Continue.
CogmemAi is the Smart Persistent Memory layer for AI agents. 95.1% on LongMemEval (April 19, 2026). 91% on LoCoMo (April 2, 2026). Built by HiFriendbot. Available as MCP server, REST API, and direct API.