A guide · ~7 min read
Fair warning before the list: we make one of these. It is in the personal-memory section and it is not the only thing there. The honest advice is to connect fewer servers than you want to, chosen around a problem you actually have.
The category with the largest gap between what ChatGPT does natively and what it could do. ChatGPT's own memory holds roughly a paragraph about you. A memory server holds years, and the model queries it rather than trying to hold it.
This is what Konshus does — https://konshus.ai/api/mcp, OAuth, read-only tokens available, and the same server answers Claude. There are also self-hosted options in this space if you would rather run it yourself; the tradeoff is that you own the uptime and the backups along with the data.
Best for: continuity across conversations and across providers.
Repository access, issue trackers, CI status. The most mature MCP category by some distance, because developers wrote the protocol and scratched their own itch first. GitHub-style servers, Linear, Sentry and similar all have working implementations.
These are also where write access earns its keep and its risk — a model opening a pull request is useful; a model force-pushing is not. Keep write confirmation on.
Best for: asking about your codebase without pasting files.
Notion, Obsidian, wikis, documentation sites. If you already keep notes somewhere structured, an MCP server over them is the shortest route to an assistant that can cite your own material instead of the open internet.
One caveat worth stating: a notes server answers "what did I write about X" well and "who am I" badly. Notes are artifacts, not a profile. That distinction is covered in using Notion as AI memory.
Best for: retrieval over material you have already written.
Database servers, warehouse connectors, analytics APIs. Powerful for asking questions in plain language against real numbers, and the category where read-only credentials are least optional. Give a model write access to production and you will eventually find out why that was a bad idea.
Best for: querying data you would otherwise write SQL for.
The rule that matters more than the list
Two servers beat ten. Every connection adds tools the model must choose between, and overlapping tools make wrong choices more likely. Connect one, use it for a week, and only add the second when you have felt the absence of it.
Related reading: how to add an MCP connector, Developer mode explained, and ChatGPT MCP vs Claude MCP.
ChatGPT remembers a paragraph about you. Konshus remembers years — and serves it over MCP to ChatGPT and Claude alike.
Meet Konshus