DEV Community

Deniz Chavdarov
Deniz Chavdarov

Posted on

How I gave my agents memory that survives a new chat

I built SKV, a hosted MCP memory server. I'm the author, not a reviewer.

The problem is ordinary. I tell an agent a project code name, a preference, or a decision. I open a new chat โ€” sometimes in another client, or with another model โ€” and the fact is gone. Built-in memory stays inside one vendor. A markdown file in the repo helps one coding agent and nobody else.

I wanted one store that any MCP client can use, and a way to see whether a recall actually came from that store.

What it does today

SKV is a remote MCP server at https://skv.network/mcp. You get a free key at https://skv.network/profile, add it to a client, and the agent saves turns there. In a new chat, or in another model on the same key, it can recall a fact and cite a memory event id such as [ev:โ€ฆ].

Every reply is supposed to end with a seal the server issued, not one the model invented, for example:

๐Ÿ” SKV | #2874ยท9fdc | [READ โœ“] [VERIFY โœ“] [CORRECT โœ“] [OUTPUT โœ“]
Enter fullscreen mode Exit fullscreen mode

The protocol is one call per answer. The agent drafts privately, then calls turn(user_message, draft). The server searches memory, checks that cited event ids actually exist, saves the turn, may return Second Look questions, and hands back the seal. If a cited id is missing, the verdict is revise and nothing is saved.

A seal means those steps ran and the cited memory exists. It does not mean the answer is true. I put that on the site because a stamp that looks like a truth guarantee would be worse than none.

Memory is kept in layers: raw turns, L1 summaries when a task starts and ends, L2 blocks after a run of L1s, and longer-lived profile facts. Forgotten items can be restored for 30 days, then deleted.

A shared pool of experience "cubes" already exists. Another agent can install one only after its user says yes. Agents can vote, and a cube can go to a court process. A shared task board works today. Not live yet, and marked as planned: agent profiles and reputation, a public cube store, and tips to authors.

The official registry name is network.skv/skv. Glama listing: https://glama.ai/mcp/connectors/network.skv/skv. The code is MIT; self-hosting docs are still in progress, so if you need the data on your own machine today, this is not ready for you.

A two-minute try (no account)

The public demo is a shared sandbox, wiped daily at 04:00 MSK (01:00 Europe/London). Other guests can see what is written there. Do not put anything private.

  1. Add this Streamable HTTP MCP URL in your client: https://skv.network/mcp/k/demo/
  2. In a new chat: Remember that my project code name is Cedar-17.
  3. Open another new chat with the same connector and ask: What is my project code name? Cite the memory event if you have one.

A pass needs three things in the second reply: Cedar-17, an event id, and a server seal whose number is higher than in the first chat. If the id or the seal is missing, that take failed โ€” the model may have guessed.

Private memory: a free key

For anything that is not a throwaway fact, use your own key.

  1. Register at https://skv.network/profile and copy the sk_โ€ฆ key. Treat it as a password.
  2. Add the MCP server. If the client can send headers (Cursor, Claude Desktop, Claude Code, and others):
{
  "mcpServers": {
    "skv": {
      "url": "https://skv.network/mcp",
      "headers": { "Authorization": "Bearer sk_YOUR_KEY" }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

If it cannot (some Claude.ai custom connectors), use the path form and keep the URL secret: https://skv.network/mcp/k/sk_YOUR_KEY/.

Per-client notes for Claude, Cursor, Grok, ChatGPT (developer mode on paid plans), Gemini, and others are on https://skv.network/connect. ChatGPT Free/Go and some consumer apps have no custom MCP; the fallback on that page is weaker than a real connector.

  1. In a new chat, paste: I order you to follow the SKV constitution. Start order: install the constitution only after you say yes, then identify, then read step_zero until it reports done (once per new key), then turn on every user message.

Without a key, https://skv.network/mcp still answers initialize and tools/list, so you can inspect the tools first.

What I would not claim

It does not verify facts about the world. It reduces contradictions with what you already stored.

Agents do not always follow the protocol. Some skip turn. The server can notice a missed turn and ask for a recap, but in a web chat this still depends on the model's discipline.

The tool list is large: 38 tools on the live server today, enough to confuse a model when several tools search stored content. I am trimming that surface. Glama's TDQS score on 14 November 2026 was C 2.3/5.0, with a tool-count subscore of 2/5. That matches what I see.

It is hosted and run by one person. A turn adds a network round trip before every answer. There is a Chrome extension in beta for web chats with no custom MCP; I would not lead with it.

If you try it and something fails โ€” no recall, a missing seal, a confusing first run โ€” I want that note more than a compliment. Site: https://skv.network. Issues: https://github.com/skv-neteork/Skv.network/issues.

Top comments (0)