Where is memory stored?

In a single local JSON file per MCP process — by default inside the working directory. Set `MEMORY_FILE_PATH` in the `env` block to pin it to a fixed path like `~/.claude/memory.json` so it persists across clients.

Can the model forget memories?

Yes. The server exposes `delete_entities`, `delete_relations`, and `delete_observations`. The agent can remove a single observation without destroying the entity it belonged to — useful for correcting wrong facts.

What is the data model — is it just key/value?

No, it is a knowledge graph: entities (nodes with a type) linked by relations, each entity carrying a list of observations. That structure matches how people actually describe what they want remembered ("Alice works at Acme, started in 2024") more naturally than flat key/value.

How big can the memory file get before it slows down?

The entire graph is read and re-written on each change, so performance degrades around a few tens of thousands of entities. For that scale, switch to a database MCP (SQLite or Postgres) and build your own structure.

Will two clients share memory if they point at the same file?

They will see the same data on startup, but each server holds the graph in-memory and overwrites the file on every mutation — concurrent writes from two processes will clobber each other. For shared memory use one dedicated client as the owner.

Memory

Name: Top MCPs — curated MCP directory
Creator: Top MCPs
License: https://creativecommons.org/licenses/by/4.0/

Persistent knowledge graph memory across AI conversations.

Score 91(?)Anthropic (modelcontextprotocol)MIT 84kVerified May 31, 2026Top MCPs for Agent Orchestration →

Install View on GitHub

Quick answer

What it does

Maintains a persistent knowledge graph of entities and relationships stored locally. The model can create, query, update, and delete memories across sessions.

Best for

Long-running agent workflows
Personal AI assistants
Project tracking
User preference storage

Not for

Single-session tasks
Structured database queries

Setup recipe

Pick your client, then follow the three steps.

Install

claude_desktop_config.jsonjson

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    }
  }
}

Paste under mcpServers. Fully quit and reopen Claude after editing.

CLI or .mcp.jsonshell

claude mcp add memory -- npx -y @modelcontextprotocol/server-memory

Run from your repo. Commit .mcp.json to share with your team.

.cursor/mcp.jsonjson

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    }
  }
}

Global path: ~/.cursor/mcp.json. Reload window after editing.

.vscode/mcp.jsonjsonc

{
  "servers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    }
  }
}

VS Code uses the "servers" key (not "mcpServers").

~/.codeium/windsurf/mcp_config.jsonjson

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    }
  }
}

Open via Cascade → hammer icon → Configure.

cline_mcp_settings.jsonjson

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    }
  }
}

Open via the Cline sidebar → MCP Servers → Edit.

~/.continue/config.jsonjson

{
  "experimental": {
    "modelContextProtocolServers": [
      {
        "transport": {
          "type": "stdio",
          "command": "npx",
          "args": [
            "-y",
            "@modelcontextprotocol/server-memory"
          ]
        }
      }
    ]
  }
}

Continue uses modelContextProtocolServers with a transport block.

~/.codex/config.tomlshell

# ~/.codex/config.toml
[mcp_servers.memory]
command = "npx"
args = [
  "-y",
  "@modelcontextprotocol/server-memory",
]

Codex uses TOML. Each server is a [mcp_servers.<name>] subtable.

~/.config/zed/settings.jsonjsonc

{
  "context_servers": {
    "memory": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "@modelcontextprotocol/server-memory"
        ]
      }
    }
  }
}

Zed calls them "context_servers". Settings live-reload on save.

ChatGPT → Apps directorynone

Memory doesn't ship a hosted HTTPS endpoint today. ChatGPT supports remote MCP servers only — to use this server in ChatGPT you'll need to deploy it to a public HTTPS URL first (e.g. via Cloudflare Workers or Vercel) or wait for an official remote build.

2
Set required secrets
No credentials required — this MCP runs over stdio without authentication.
3
Try a minimum working prompt
Teach the agent persistent facts about a project
Remember these facts about my current project: (1) it is a CRM called Atlas, (2) stack is TypeScript + Next.js + Postgres, (3) I prefer minimal dependencies and terse code. Tomorrow, confirm you remember all three before writing any code for this repo.
Tested with: Claude Desktop, Cursor.

Tools & permissions

Tool	Description	Args	Side effects
`create_entities`	Create named entities with types and observations.	entities: object[]	Write
`create_relations`	Create typed relations between existing entities.	relations: object[]	Write
`add_observations`	Append observations to existing entities.	observations: object[]	Write
`delete_entities`	Delete entities and their relations.	entityNames: string[]	Write
`delete_observations`	Remove specific observations from entities.	deletions: object[]	Write
`delete_relations`	Remove relations between entities.	relations: object[]	Write
`read_graph`	Return the full knowledge graph.	—	Read
`search_nodes`	Search for entities by query.	query: string	Read
`open_nodes`	Return specific entities by name.	names: string[]	Read

Security & scope

Access scope: Read + write
Sandbox: Persists a knowledge graph as a single JSON file on disk (default `~/.mcp/memory/memory.json`, override via `MEMORY_FILE_PATH`). No network access.
Gotchas: The memory file is plain JSON — if you store sensitive information the server does not encrypt it.
No concurrency control — avoid two clients writing the same file simultaneously.

Agent prompt pack

— copy into Claude, Cursor, or ChatGPT.

Paste into Claude, Cursor, or ChatGPT. Edit the [brackets] before sending.

Recommend the best MCP servers for [task: e.g. agent orchestration work] in [client: Claude].

Constraints:
- Prefer tools that are [official | open-source | read-only] — pick what matters for my use case.
- Exclude MCPs that require [e.g. a paid plan, OAuth-only flows, remote-only transport].
- Return at most 3 picks, ranked.

For each pick include:
1. One-sentence rationale.
2. The ready-to-paste install snippet for my client.
3. Any required secrets I need to create before installing.

Cross-check the top-mcps.com listing: https://top-mcps.com/top-mcps-for-agent-orchestration

Compare Memory against a real alternative. Swap the second MCP in [brackets] if you want a different match.

Compare Memory MCP vs [SQLite MCP] for the following job: [describe the job, e.g. "let an agent create GitHub issues on bug triage"].

Judge them on:
- Setup time and complexity (what a new user hits first).
- Auth model (none / API key / OAuth 2.1) and credential risk.
- Transport (stdio / Streamable HTTP / SSE) and where the server runs.
- Required secrets and the blast radius if they leak.
- Operational risk in an unattended agent loop.
- Which one is "good enough" for a weekend prototype vs. production.

End with one sentence: which should I pick for my scenario, which is: [my scenario].

References:
- https://top-mcps.com/mcp/memory
- top-mcps.com listing for SQLite

Asks the agent to install and verify. Works inside Claude Code, Cursor Agent, Codex CLI.

Install the Memory MCP server for my [client: Claude] at the default config path for that client.

Use the exact install snippet published at https://top-mcps.com/mcp/memory (fetch https://top-mcps.com/mcp/memory.json for the canonical server.json if you can read URLs).

Before finishing:
1. Create the required secrets (no secrets) and put them in the appropriate env block — do not hard-code them.
2. Restart or reload the client so it picks up the new server.
3. Verify the server is connected (green / running state) and at least one tool is listed.
4. If anything fails, read the client's MCP logs and report the exact error — do not silently retry.

Confirm when done and list the tools the server now exposes.

Frequently asked questions

What changed

— 2 updates tracked.

May 31, 2026
Refreshed install snippets and fact sheet; verified for 2026.
Nov 25, 2024
Initial directory listing.

More Agent Orchestration MCPs

Other tools in the same category worth evaluating.

Sequential Thinking

Structured step-by-step reasoning for complex problem solving.

reasoningthinkingagentsplanning

2 minLow

Claude Task Master

AI-driven task management and decomposition for long-running agent projects.

agenttasksplanningpersistence

3 minLow

Context Forge

Capture context and decisions across sessions — durable agent memory in markdown.

agentmemorycontextpersistence