← Tutti gli articoli

The best memory MCP servers in 2026, by use case

There is no best memory server. There is a best one for local-only, one for self-hosting, one for graphs, and one for a memory that follows you across assistants.

Pubblicato il 23 settembre 2026 · 7 min read

A year ago “memory for your AI” meant one thing: a vendor feature inside one app. Then the Model Context Protocol gave assistants a standard way to call outside tools, and memory became something you could plug in — and swap. Today the search for a memory MCP server returns a dozen projects that do not compete on the same axis at all.

This is a guide to eight of them, sorted by what they are actually for. Every fact was read on the vendor's own site or repository on 22 September 2026 and is the same one used in our full comparison table. Prices move; follow the link before deciding.

The one question that eliminates half the list

Before features: does the server run on your machine (stdio) or somewhere reachable (HTTP plus OAuth)? A local server is fast, private and free, and invisible to Claude on your phone, to ChatGPT, or to any browser client. A remote server is a URL any client can sign in to, from any device — and someone operates it, which means you should know who and where.

Neither is better. One memory on one laptop is a notebook; one memory across every assistant is the actual promise. Decide which you want, and most of the list falls away on its own.

Knowledge Graph Memory Server — the official reference

The memory server in the official MCP servers repository is where most people start. It speaks stdio, writes entities and relations to a JSONL file on disk, and does substring matching — no vectors, no ranking. It is free and dependable.

Choose it if you want to understand what an MCP memory even is, on one machine, in ten minutes. It is local only, so it will never appear on a phone or in a browser, and it will not find a note phrased differently from how you wrote it.

mcp-memory-service — self-run, with real retrieval

mcp-memory-service is an Apache 2.0 server you host yourself, local or on your own box. It combines a local vector index with BM25, and it supports OAuth 2.0 with dynamic client registration — the right architecture for remote access, with you as the operator. Free.

Choose it if you want hybrid search without a vendor and are comfortable deploying and exposing a service. You get everything below except the part where someone else keeps it running.

claude-mem — local-first, built for Claude Code

claude-mem is an Apache 2.0 engine that runs local-first with SQLite FTS5 plus vector search; remote access is reserved for its paid cloud tiers. The engine is free, the cloud plans start at $20, with a $333 per-seat tier.

Choose it if your assistant is Claude Code on one machine and you want it to remember sessions with zero setup. Its centre of gravity is the terminal; a memory shared with ChatGPT in a browser is not what it is for.

basic-memory — notes you own, in Markdown

basic-memory is AGPL-3.0, self-hosted for free, with cloud plans at $15 and $30 per seat for remote access. Retrieval is full-text plus vector, with an optional reranker.

Choose it if you want the memory to be plain files you can open in any editor, and copyleft licensing is a feature to you rather than a constraint.

Hindsight — when the question spans several memories

Hindsight is MIT, self-hosted free with a usage-based cloud, and exposes one MCP endpoint per memory bank. Its retrieval stacks vector, BM25, a graph layer and temporal reasoning — the multi-hop kind of question (“what changed between the two clients I met in March?”) that ranked search alone answers badly.

Choose it if your memories are entities with relations and dates, not notes. You pay for that in operations: it is yours to run.

Mem0 — the most cited, cloud or self-hosted

Mem0 is Apache 2.0, available as a hosted MCP endpoint at mcp.mem0.ai or self-hosted. Retrieval is vector plus keyword with reranking, and a graph layer on the Pro tier. Free tier, then $19 and $249 a month. GDPR compliance, SOC 2 Type 1 and HIPAA are claimed; the hosting region and an EU residency option are not something we could confirm from public pages, so we make no claim either way. OpenMemory, its former local MCP product, was removed from the monorepo in July 2026.

Choose it if you are building an agent and want the library everyone else's tutorial uses, with the option to run it yourself later. It is developer-shaped: its natural home is a codebase, not a settings screen.

Supermemory — hosted, hybrid, open

Supermemory is MIT-licensed, hosted at mcp.supermemory.ai or self-hosted, with semantic and hybrid search over a vector-graph store. Free tier, then $19, $100 and $399. GDPR compliance is claimed; hosting region, like Mem0, is not confirmable from public pages.

Choose it if you want a hosted endpoint today and the freedom to read or fork the code tomorrow. That combination is rare on this list.

Pryzm Memory — one hosted memory, every assistant

Pryzm is the remote kind only: one hosted endpoint, OAuth 2.0 sign-in, no local install, shared by every client you connect. It runs on servers in Germany under EU jurisdiction, encrypted at rest, with one PostgreSQL schema and role per customer enforced by the database itself. Retrieval is hybrid — lexical and vector fused by reciprocal rank, then re-scored by a multilingual cross-encoder — across eight languages, tuned for French as much as English. Free for 100 memories, then €9.99 and €19 a month.

Choose it if the point is one memory that follows you from Claude to ChatGPT to your phone, and you would rather know the jurisdiction than run a server.

When Pryzm is the wrong choice

One line each

The dated, source-linked version of all of this — with hosting, licence, retrieval and entry price side by side — is the comparison table. If something there is wrong about your product, the page says how to get it corrected.

Connect a memory in two minutes