What is an MCP memory server? The protocol, the tools, and the tokens
Summary
A vendor of an AI memory product published a video and a written companion explaining what an MCP memory server actually is, opening with a disclosure that it builds one. The walkthrough uses only the public reference memory server and the Model Context Protocol specification itself, with code excerpts as selectable text and each source linked. The definition it lands on is narrow: a small program that exposes a handful of tools over MCP and writes what it is told to disk so the content survives the end of a conversation.
The sharper point is about what the protocol does not cover. MCP standardises tool discovery, tool calls, and the shape of the response. It does not standardise memory. The TypeScript schema mentions memory once, in a comment about a boolean hint, and searching it for persistence, storage, retrieval, vector, or session returns nothing. The July 2026 revision describes MCP as a stateless protocol and the tools page states there is no protocol-level session, so what a server stores and what it returns is that server's own design.
Demonstration by Mnemoverse: AI Memory & Agents, embedded from YouTube. AIOTruth did not reproduce the test.
Why it matters
Anyone publishing about MCP memory is writing into a space where the shared vocabulary does not exist at the protocol layer, which means models answering questions about persistence, retrieval, or sessions are pulling from vendor documentation rather than from a common standard. If your product stores context for an agent, the retrieval behaviour, the eviction rules, and the token cost of a recall are yours to describe, and nothing in the specification will describe them for you. That makes the written page carrying linked, checkable excerpts the asset a model can cite, and it makes vague claims about being MCP compatible useless as differentiation, because compatibility says nothing about the part a buyer is actually choosing between.
Source
- mnemoverse.com/docs/library/what-is-an-mcp-memory-server the item itself, published by Mnemoverse: AI Memory & Agents, created by Mnemoverse: AI Memory & Agents
- github.com/modelcontextprotocol/modelcontextprotocol/blob/main/schema/2026-07-28/schema.ts Public source code or repository
- modelcontextprotocol.io/specification/2026-07-28 Published technical standard
- modelcontextprotocol.io/specification/2026-07-28/architecture Published technical standard
- youtube.com/watch?v=MQDFOmsMOY0where AIOTruth found it
What was checked
- canonical URL taken from the page itself
- read the item page (200)
- https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/schema/2026-07-28/schema.ts: supports the item, tier 6 (Public source code or repository)
- https://github.com/modelcontextprotocol/servers/tree/7b1170d/src/memory: does not mention the item, tier 6 (Public source code or repository)
- https://npmjs.com/package/@mnemoverse/mcp-memory-server: unreachable (403)
- https://mnemoverse.com/docs/library/what-is-an-mcp-memory-server: supports the item, tier 2 (Reproducible first-party demonstration)
- https://modelcontextprotocol.io/specification/2026-07-28: supports the item, tier 3 (Published technical standard)
- https://modelcontextprotocol.io/specification/2026-07-28/architecture: supports the item, tier 3 (Published technical standard)
Limitations
- 1 linked source did not respond when checked and could not be used as evidence.
- The demonstration was not independently reproduced by AIOTruth.
AIOTruth judgment
Evidence status is Confirmed, and the evidence dimension scores 10 because the central claims point at named artefacts anyone can open: the schema file at the 2026-07-28 path, the specification page, and the architecture page, all of which were fetched and confirmed to discuss this item. Usefulness at 8 reflects that the distinction between what MCP standardises and what a server designs is the distinction a reader needs before comparing products. Originality at 7 credits the framing rather than the underlying facts, which are public. Relevance at 6 is the constraint: this is protocol mechanics, adjacent to discovery rather than about it. The vendor disclosure appears in the first line and no product of theirs is shown, which is why the vendor origin does not pull the rating down. Two limitations stand: one linked source did not respond when checked, and the demonstration was not independently reproduced by AIOTruth.
How this score was calculated
| Dimension | Weight | Score | What produced it |
|---|---|---|---|
| AIO relevance | 30% | 6 | rag, model context protocol; 4 scope question(s) matched |
| Usefulness | 30% | 8 | 2 artifact(s), 6 actionable marker(s), 0 measured figure(s) |
| Evidence | 25% | 10 | Reproducible first-party demonstration; 4 verified source(s) across 3 domain(s) |
| Originality | 15% | 7 | no original testing found; closest archive match 0 |
aioRelevance x 0.30 + usefulness x 0.30 + evidence x 0.25 + originality x 0.15. The rubric is published in full on the Editorial Method page. Scoring is deterministic: the same item scores the same on every run.
Read the source Open the source
Found something wrong here? AIOTruth corrects material errors openly. Have something we should review? Submit a find.