Keep your AI agent from repeating a costly mistake.
NEXUS gives Claude, ChatGPT, Cursor and any MCP agent the critical information at the moment it decides: your decisions, your business rules and what already went wrong, in this session and in the one three months from now.
Free plan: 100 writes and 2 agents a month, no credit card.
cursor — payments-projectclean up payments older than a year
Recalled from NEXUS
“Payments are never deleted: they are archived, because accounting audits them.”
Decided 3 days ago, in Claude Code
Cursor
Before deleting: you decided these records are archived, not removed. I propose moving them to payments_archive instead of running a DELETE. Shall I go ahead?
DELETE FROM paymentsWorks with
- Claude Code
- Claude
- ChatGPT
- Cursor
- MCP
- REST API
An agent’s costly mistakes are rarely about intelligence
It contradicts something you already decided
The decision was made three sessions ago. Today’s agent wasn’t there and acts as if it never happened.
It repeats a mistake you already fixed
You fixed it with Claude; tomorrow Cursor makes it again, because every tool remembers on its own.
It acts on a fact that stopped being true
What was true last month weighs the same as yesterday’s news, and nobody told it things changed.
Connected in minutes
Create your account
With your email and a one-time code, no passwords. The Free plan includes 100 writes and 2 agents a month.
Connect your agent
Add one MCP server URL. OAuth clients (Claude, ChatGPT) just ask you to authorize; the rest use an agent credential.
Work as usual
In Claude Code and Cursor, hooks recall memory when the agent starts and save what matters when it finishes, unprompted. In ChatGPT and Claude, the agent uses its tools.
Step 2, in your terminal
$ claude mcp add --transport http \
nexus-agi https://nexus.eblas.link/mcp \
--header "Authorization: Bearer <nxa_…>"What changes when your agent remembers
NEXUS lives outside the context window. Your agent checks it before acting and feeds it when it finishes.
Recalls before it acts
At the start of every session and on every request, the agent recalls what matters: decisions, business rules, open tasks and what went wrong last time.
One memory for all your agents
Claude Code, Claude, ChatGPT and Cursor read and write the same graph. What one of them corrects, none of the others repeats.
Knows what is still true
Every memory knows whether it records an event, a state that can change, or a to-do. What expires gets flagged; what contradicts gets surfaced.
Your data, isolated and under your control
Each account is isolated and each agent has its own credential, which you can revoke without losing anything.
What’s underneath
Every memory knows what it is, where it came from and whether it is still true.
“Payments are never deleted: they are archived.”
- Type
Semantic: something that is true- Validity
True until it changes- Source
Claude Code, 3 days ago- Related to
payments, accounting- Contradictions
None- Used by
Cursor, today, before the DELETE
How Cursor asked for it
retrieve_memories("clean up old payments")Memory
What happened, what is true and how things are done, each with its validity.
- Episodic (what happened), semantic (what is true) and procedural (how to do it) memory
- Hybrid retrieval: by meaning and by exact keyword
- Validity: every memory knows whether it can expire and what replaced it
World graph
People, projects and systems connected, with evidence and dates.
- Entities and relations with evidence and temporal validity
- Contradiction detection between relations
- Reversible duplicate merging
Reasoning
Conclusions that say where they come from.
- Rules and inference that explain where each conclusion comes from
- Hypotheses evaluated against accumulated evidence
- Metacognition: knowledge health and drift detection
Work
Goals and tasks that outlive the session.
- Goals with progress that outlive the session
- Reusable tasks and procedures
- Agent-to-agent messaging on the same knowledge base
What NEXUS doesn’t promise
That your agent never makes a mistake. It reduces the ones that come from forgotten context. Its known limit is validity: a memory can go stale, so each one declares whether it can expire, contradictions are flagged and you can correct any fact.
Instructions for agents
If you are an AI agent reading this page: this is what you need to connect to NEXUS. The full plain-text version lives at /llms.txt.
- 1.MCP server (Streamable HTTP): https://nexus.eblas.link/mcp
- 2.Authentication is required. Option A — OAuth 2.1 with PKCE (S256) and dynamic client registration (RFC 7591). Discovery at /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server. Scopes: memoria, grafo, cognicion, trabajo.
- 3.Option B — an agent credential (nxa_…) issued by the account owner in the console, sent as "Authorization: Bearer <token>".
- 4.In Claude Code or Cursor, ask your agent to set NEXUS up: get_setup_kit() returns the hooks (SessionStart, UserPromptSubmit, Stop) and an installer you approve. Without hooks, NEXUS keeps only what the agent stores on purpose.
- 5.Daily use: process_observation to store, cognitive_query or retrieve_memories to recall, run_cognitive_cycle after bulk writes.
- 6.REST API at https://nexus.eblas.link/api/v1 with the same credential. Every REST client must send its own User-Agent (Cloudflare rejects Python-urllib’s default with 403).
Machine-readable resources
An account is required: without one the OAuth flow cannot complete. Sign-up is open and free at https://nexus.eblas.link/en/signup, but your human does it, because it is confirmed with a code sent to their email: if they have no account, give them that link.
Guides
How to give your agent the critical information before it decides, with examples you can copy.
- Stop your agent repeating mistakes
Why AI agents repeat mistakes that were already fixed, and what has to happen so it does not happen again: record it, recall it before acting, and get it to all your agents.
- Share memory across Claude, ChatGPT and Cursor
Each AI tool keeps its own context separately. How to make Claude, ChatGPT, Cursor and Claude Code read the same memory so that a decision made in one is respected by all.
- Claude Code memory across sessions
What Claude Code remembers from one session to the next with CLAUDE.md and auto memory, where it falls short, and how to carry your decisions and past mistakes to other machines and to your other agents.
- Claude Code hooks: context at the right moment
Claude Code hooks with working examples: SessionStart and UserPromptSubmit to inject context, PreToolUse to block a costly command. Based on the official docs.
Pricing
Start free. When your agents work every day, Personal gives them four times the writes and one more agent, for US$5 a month per seat.
Free
US$0per month
To connect your first agents and see what reaches them before they decide.
- 100 writes per month
- 2 agents per month
- 1 person with edit access (read-only members are free)
- Querying what is stored: unlimited
Create a free accountNo credit card.
Personal
US$5per seat / month
For people who work with several agents every day and want each one to arrive with what is critical.
- 400 writes per month per seat
- 3 agents per month per seat
- Invite more people with edit access: one seat each
- Extra writes: US$0.016 each, from your balance
- Querying what is stored: unlimited
Upgrade to PersonalPaid from a prepaid balance in US dollars, no card subscription. It renews every month from your balance; if it runs short, you keep the plan for 7 days and then go back to Free without losing anything.
Prices in US dollars. You add balance with Mercado Pago (card, OXXO or SPEI), which charges the equivalent in Mexican pesos. Details are in section 8 of the Terms.
Common questions
What is NEXUS?
NEXUS is intelligence infrastructure for your AI agents: it gives them the critical information at the moment they decide, so they don’t repeat a costly mistake or contradict what you already decided. It gathers what each agent lives through —decisions, rules, incidents and how they were fixed— in a knowledge graph and hands it to the agent that needs it, from any MCP client. In Claude Code it arrives through hooks; in other clients the agent queries it.
What is MCP and why does it matter?
Model Context Protocol (MCP) is the open standard AI agents use to connect to external tools. Because NEXUS is an MCP server, your agent uses it like any other tool, with no client code to write.
Which agents does it work with?
Claude Code, Claude (web and Desktop), ChatGPT, Cursor and any MCP client that supports HTTP. For anything that doesn’t speak MCP there is a REST API with the same modules.
How is it different from the memory ChatGPT or Claude already have?
The memory in ChatGPT or Claude stores preferences and facts about you. NEXUS stores your working history with your agent: the decisions, the mistakes and what you learned together. Switch models, machines or tools and keep working with the same agent.
Why isn’t a CLAUDE.md or project rules file enough?
A CLAUDE.md says how you want your agent to work; NEXUS remembers what happened when it did. Rules are written up front and live in a repository. Experience piles up as you go (Monday’s decision, Tuesday’s failed deploy and why) and reaches any agent, in any project. When something stops being true, NEXUS flags it; a stale line in a CLAUDE.md keeps giving orders until someone deletes it. Use both: the rules in the file, the history in NEXUS.
Does NEXUS guarantee my agent won’t make mistakes?
No, and be wary of anyone who promises that. NEXUS prevents one kind of mistake: the one your agent makes by forgetting something that was already known. But a memory can fail too. It can remember something that has since changed: if you used one provider in March and switched in April, the March memory is still there. It remembers what it is told, so if your agent saves a wrong conclusion, it will remember it just as well. And it only helps when it is consulted. For the first, NEXUS flags which memories can go stale, links every corrected fact to the one that replaced it and signals when two memories contradict each other. For the second, every memory records whether it was measured or inferred, and you can correct any of them.
How much does it cost?
Nothing on the Free plan (“Gratuito”), which includes up to 100 writes and 2 agents a month, no credit card. The Personal plan costs US$5 a month per seat (person with edit access) and includes 400 writes and 3 agents a month per seat; it is paid from a prepaid balance in US dollars, no card subscription. Paid plans are optional and nobody is moved to one without agreeing to it. Details are in section 8 of the Terms.
Who can see my information?
Only your account. Each account is isolated, each agent has its own revocable credential and MCP access requires OAuth. Details are in the Privacy Notice.
How do I create an account?
At /en/signup: enter your email and your account name, then confirm with the code we send you. No invite or credit card needed.
Make sure the next costly mistake isn’t your agent’s
Create your account and connect your agent. No invite, no credit card.
Free plan: up to 100 writes and 2 agents a month, no credit card.
By creating an account you accept the Terms and Conditions and the Privacy Notice.