An MCP connection is a way for an assistant to call tools; it is not proof that the underlying memory runs locally, that every client can connect, or that the system automatically remembers every conversation. The memory server, model, client and storage boundary all matter.
NeuroAIgent publishes this guide and makes Central Brain. This official-documentation review is dated October 17, 2026. Our Central Brain recommendation reflects the documented local document-and-memory workflow. It is not an independent lab ranking or a head-to-head performance benchmark.
The five criteria that change the decision
First, define the input: selected project files, curated notes, captured work activity, user interactions or conversations with a stateful agent. Second, identify the deployment: a local desktop process, a self-hosted service or a hosted endpoint. Third, identify what persists and whether a person can inspect or edit it.
Fourth, check the actual client and transport, including account, workspace policy and tool permissions. Fifth, define an evidence check: can the assistant retrieve the right source and carry a corrected decision into a new session? A long integration list cannot answer that last question for your files.
| Product | Documented delivery | Memory structure |
|---|---|---|
| Central Brain | Local desktop app exposing MCP tools | Selected documents plus readable project notes |
| Basic Memory | Local MCP or hosted Cloud | Linked Markdown knowledge and note editing |
| Mem0 OpenMemory | Official guide describes local MCP service | Cross-client memory with a configurable engine |
| Supermemory | Hosted MCP; separate self-hosted API option | Shared spaces, documents, memories and profiles |
| Zep | Hosted Context MCP and client plugin | User-scoped context graph in a Zep project |
| Letta | Hosted MCP to create and message agents | Delegate to stateful agents with persistent memory |
Central Brain: selected project evidence and readable continuity
Central Brain is our recommended starting point for a consultant, researcher, developer or small team with an existing folder of project evidence. The Windows and Apple silicon Mac desktop product prepares PDF, DOCX, TXT, Markdown, CSV and XLSX files, creates a local index with a bundled embedding model, and serves search and memory tools over MCP. Confirm the available installer for your platform before buying.
Its project record consists of five note types and a project brief in readable local files. Exact and semantic search complement that record; result traces include source path, prepared-file path, page and chunk where available. Its eight tools separate read-only file access from memory writes. A RESUME block on connection helps the next session pick up the saved project.
Central Brain combines these steps in a desktop workflow with storage you choose. Licensing checks occur online every 14 days; Team and Enterprise can obtain a no-online-check build. The connected AI receives retrieved context, and optional cloud connections are enabled deliberately. Your project stays organized around its saved brief, notes and source files.
Basic Memory: a linked Markdown knowledge base
Basic Memory documents local and hosted modes, a Markdown record with observations and relations, and MCP note operations. These describe a note-graph architecture. Central Brain organizes the project around selected documents, five checkpoint types and a readable brief.
Choose Central Brain when your starting point is a mixed document folder and you want preparation, source-traceable search and compact project memory in the same desktop app. Your evidence and saved decisions become one repeatable workflow across supported AI sessions.
Mem0 OpenMemory: distinguish the MCP service from the API
Mem0 publishes an OpenMemory MCP guide describing local cross-client memory and a management interface. Its broader product family also includes a managed memory platform and an open-source engine. Do not infer the deployment or provider configuration of one from the name of another.
For a developer, the important questions are which current distribution you will run, which models and embedding providers it uses, and who maintains the service. Central Brain is designed for the buyer who wants a packaged local project workflow instead of assembling those memory-engine components. Neither a local database nor an MCP label alone proves that no content reaches an external model.
Supermemory: hosted shared spaces, plus a separate local API
Supermemory documents a hosted MCP endpoint with OAuth, spaces, document operations and profile context. Its repository describes the corresponding tools and MCP Apps. This is a hosted context-service architecture.
Supermemory also documents a separate self-hosted API server with local embeddings and an offline text-processing configuration. Its current table lists MCP for the hosted platform rather than that local server; some file paths use external services. Central Brain brings local document preparation and local MCP together in its own desktop workflow.
Zep: user-scoped context through a hosted MCP connection
Zep v4 documents a Context plugin that installs both a hosted MCP connection and a skill guiding when to search or save memory. Administrators configure a project; people sign in and access their own user graph. The tools read and write context in that project.
That describes an administered context service with user identities and graphs. Choose Central Brain for a locally prepared evidence folder and an inspectable project record on storage you choose. Its desktop workflow keeps preparation, retrieval and saved decisions together.
Letta: reach a stateful agent over MCP
Letta documents a hosted MCP server for listing, creating and messaging stateful agents and retrieving their replies. That exposes an agent-delegation interface, with memory attached to those agents.
Central Brain serves project context directly to the supported assistant you already use. You can search an approved PDF, inspect its source and retrieve a saved project decision through the same project connection. That is the practical workflow this guide recommends.
A useful evaluation takes one real project
Use an approved, representative sample with known answers. Ask one factual question that requires a source, one question involving a changed decision, and one question whose answer is absent. Record whether the tool returns evidence, respects the correction and admits the missing information. Resume in another session and, if needed, another supported client.
Inspect the saved record and the outgoing context boundary. Check backup, removal, project selection and who can access the storage. These are acceptance criteria rather than published benchmark results; no performance number is implied. If the job is local document preparation and portable project continuity, choose a Central Brain plan and start with that workflow.
Common questions
What is the best local MCP memory server for project documents?
For a packaged desktop workflow that prepares supported local documents and shares readable project memory, our recommendation is Central Brain. This is our product-fit assessment as its publisher, not an independent benchmark result.
Are all six products local MCP servers?
No. They expose different deployment and MCP models. For example, Zep and Letta document hosted endpoints; Supermemory documents hosted MCP separately from its local API server. Check the exact offering, not just the brand name.
Does this roundup rank retrieval accuracy?
No. It compares documented workflows and deployment choices. A fair accuracy comparison would need the same dataset, questions, configurations, scoring and test conditions.