Shared agent memory without a server-maintenance side project

You can give your team's agents shared memory without running a server. A managed service handles installs, upgrades and configuration, so setup is a sign-in and a connection instead of a project. Nysa works this way: setup takes under 5 minutes, and Nysa runs and updates the brain for you.

The rest of this guide explains what the maintenance work is, why it tends to grow, and what the managed alternative looks like.

The side project nobody planned

Most teams do not set out to run infrastructure. They want their agents to remember things. Then they pick a self-hosted memory tool, and a few evenings later someone owns:

• A server and a database to run and back up.
• API keys for the models the tool calls.
• Credentials and an OAuth app to read email or calendars.
• A way to get meeting transcripts into the system.
• Upgrades, every time the tool ships a new version.

None of this is hard on its own. The trouble is that it is permanent. The person who set it up becomes the person who fixes it, usually at the worst moment.

What people report going wrong

Public issue trackers for self-hosted memory tools are a useful window into this. Across them, similar kinds of problems come up. These are common reports about self-hosted memory layers, not claims about any one project's current state, and many get fixed.

• Fresh installs that stumble. For example, a report where the first sync of a new install was blocked^[1] by files the install itself created.
• Upgrades that need hands. One report describes an automatic upgrade that stopped unattended work and needed manual intervention.^[2]
• Configuration that is ignored. Another describes a documented setting being silently replaced by a default.^[3]
• Proxy and deployment workarounds. Normal hosting setups sometimes need fragile patches before they work.

The pattern is not that the tools are bad. Memory layers have many moving parts: a database, a model, connectors, background jobs. Each part can drift, and when you host it, you are the one who notices.

What managed memory changes

With a managed service, the moving parts still exist, but they are someone else's job. For Nysa, that looks like this:

1. Sign in with Google. Your brain starts filling from meetings, Gmail and Google Calendar.
2. Invite your team. Teammates join the same workspace.
3. Connect your agents. Claude, ChatGPT, Cursor, Codex and other MCP clients connect to one server at https://mcp.usenysa.com/mcp, signing in with OAuth.

There is no database to provision, no OAuth app to create, no notetaker to wire in. Nysa's own notetaker joins the calls you choose, as a visible participant. Gmail history is imported when you connect, and new mail keeps flowing in.

For a step-by-step view, see connect Claude and connect ChatGPT, or read how Nysa turns meetings into facts.

What stays your job

Managed does not mean zero effort. Be honest about what you still own:

• Choosing which calls to record. The notetaker joins the calls you pick.
• Inviting teammates and deciding who connects which agent.
• Judging the answers. Agents read and write the brain, so it is worth glancing at the sources behind important facts.

This is a few minutes of attention, not a standing maintenance task. Nysa works with Google accounts today: Gmail and Google Calendar, personal or Google Workspace.

When self-hosting is the right trade

Self-hosting suits teams that must run everything on their own infrastructure, or that want to rewrite how memory works. If that is you, read self-hosted vs managed company brain for a way to decide. For most small teams whose real goal is "our agents should know what we know", maintenance adds nothing. Nysa gives them what a self-built setup has to assemble: a notetaker built in, 180 days of email and calendar history on day one, facts with sources, and access scoped to each teammate.

A quick check for any option

Before you commit, ask the person who would run it:

• Who gets paged when ingestion stops?
• How do upgrades happen, and what breaks if they fail?
• Where do keys and credentials live?
• How long until a new teammate's agent is connected?

If the honest answer to the first question is "me", count that time. The next guide, what agent memory actually costs, breaks that cost down.

FAQ