Memory Infrastructure for AI Agents

Keep agent context
across sessions and tools.

Nexus stores decisions, facts, and work history so the next agent, session, or tool can continue without a full re-brief.

You get a hosted memory API, an MCP server, AMCP reference endpoints, and portable export/import in one product. Start local, then expand to personal cloud memory or shared team memory when you need it.

Use hosted remember, recall, sessions, and export/import endpoints without building your own memory layer.

Connect supported coding-agent clients through one MCP installation path.

Scale from Nexus Free on one device to Nexus Pro and Nunchi Team without changing the core memory model.

Open Product ->Get API key ->Read the docs

Install Nexus MCPHosted setup needs a Nexus API key

npx @nunchiai/nexus-mcp@latest init --client codex --key <YOUR_NEXUS_API_KEY> --yes

AMCP reference server · MCP-native · REST API · Export / import portable

Why agents need a memory layer

No memory of past conversations

The user repeats the same context. The agent answers like it has never met them before.

No context across sessions

Yesterday's decision cannot shape today's answer. Experience does not accumulate.

Hard to recover prior decisions and rules

As conversations grow longer, early decisions and constraints disappear and the same reasoning gets repeated.

Nexus handles capture, structure, and recall.

It stores agent decisions, normalizes them with scope and retention, then returns the right context when the next session or tool needs to continue the work.

Capture

MCP and API entry points bring memory into the agent loop with minimal friction.

Structure

Context is normalized into portable memory objects with scope, origin, and retention.

Recall

Relevant context returns when the next agent, session, or tool needs to continue the work.

What ships today

Nexus is memory infrastructure for agent runtimes, not a note-taking UI. It keeps the same memory operations and portability model across APIs, MCP clients, and AMCP-compatible runtimes.

How it works

const nexus = new NexusClient({
  apiKey: process.env.NEXUS_API_KEY,
  agentName: "your-agent",
});

await nexus.remember("Auth flow uses PKCE", "decision", {
  scope: { kind: "project", id: "acme-web" },
});

const context = await nexus.recall("auth decisions", {
  scope: { kind: "project", id: "acme-web" },
});

AMCP-native · MCP-native · Hosted API · Export / import portability · Works with any agent framework

Benchmark Result

83.2% on LongMemEval_S, measured on the production memory pipeline.

This was not a leaderboard-only prototype. We measured the same Synapsis memory pipeline that handles real API traffic, then published the remaining failure cases and limitations.

Read the benchmarking note ->

Read the note for the production setup, failure cases, and why we think honest memory benchmarks matter more than vanity scores.

Core capabilities

Store and retrieve decisions, facts, and work history across sessions.

Use the same memory through hosted API, MCP server, and AMCP reference surfaces.

Start with local memory, then expand to personal cloud memory or shared team memory.

Connect durable memory
to your agent workflow.

Open Product ->Get API key ->Read the docs Talk to us