DEV Community

Coding-Dev-Tools
Coding-Dev-Tools

Posted on

An MCP memory server should answer three questions before it answers yours

Most "agent memory" tools start with storage: embeddings, a vector table, a transcript index. Storage is the easy part. The failure that actually breaks a coding agent is quieter: it retrieves a fact that was true last month, presents it as current, and cannot show where it came from.

An MCP memory server should answer three questions before it answers the user's question.

1. What is this claim scoped to?

A convention for one repository is not a convention for every repository. A decision for one service is not a decision for the whole organization. Scope is not metadata you add later; it is part of the claim. A useful memory record says "this applies to this repo, this service, or this user" and refuses to leak across those boundaries by default.

2. When did it stop being true?

Most systems append. Few can supersede. If the team changed from Jest to Vitest, the old note is still in the index and still looks relevant. A memory server needs a time window and a replacement chain. When the agent asks what the test runner is, it should get the current answer, not a confident answer from three migrations ago.

A useful test: ask the same question against two dates. The answer should be different, and the system should be able to explain why.

3. What record supports it?

If the memory cannot point to the decision, commit, issue, or conversation that supported the answer, the agent cannot tell a current convention from a stale one. Provenance is what makes the difference between a memory and a rumor. It also gives the user a way to correct the record instead of arguing with a black box.

The server should be able to abstain

A retriever that always returns five passages has no way to say "I do not have this." That is not a memory system; it is a confident randomizer. The useful behavior is a bounded answer with citations, or an explicit abstention when the record is missing or conflicting.

This is the checklist I use before an agent touches production:

  1. What is the scope?
  2. What is the time window?
  3. What supports the claim?
  4. What replaces it when it changes?
  5. Can the system return nothing?

The full 12-question checklist is here: https://engraphis.com/agent-memory-checklist

I maintain Engraphis, a local-first memory engine for coding agents. It stores scoped claims, keeps temporal history, cites the supporting record, and exposes the store over MCP so Claude Code, Cursor, and Codex can use the same memory. If you are evaluating an MCP memory server, start with the questions above before comparing feature lists.

Top comments (0)