--- title: "Best Memory MCP Servers for Claude Code & Cursor" description: "Compare 8 memory MCP servers for Claude Code and Cursor in 2026, including past.dev, mem0, Supermemory and Zep, with setup steps, pricing and BEAM results." canonical: https://past.dev/blog/best-memory-mcp-servers-claude-code-cursor date: 2026-11-19 last-updated: 2026-11-19 category: Engineering authors: The past.dev team --- # Best Memory MCP Servers for Claude Code & Cursor The best memory MCP server for Claude Code and Cursor keeps one memory that both tools can read, returns the current version of each decision with its source, and recalls accurately as history grows. past.dev fits that brief: its remote MCP servers need nothing installed, its open-source plugins add recall to every prompt in Claude Code, Codex, Cursor, GitHub Copilot, Hermes Agent and OpenClaw, and it ranks first on BEAM at every history size among published results (self-reported, and the harness is public). claude-mem, Supermemory, mem0, Zep, Graphiti, Cognee and the MCP reference server fit narrower setups. Each entry below covers what the server is, how it runs, its licence, how to connect it to Claude Code and Cursor, its published results and when to choose it. ## Why Do Claude Code and Cursor Need a Memory MCP Server? Claude Code and Cursor start each new session without the context of earlier ones. Developers re-explain project structure, conventions, architectural decisions and preferences, and on larger codebases or teams that repetition introduces inconsistency. ### Problems Developers Encounter Without Persistent Memory - **Session reset:** decisions made last week are invisible in today's session. - **Repeated re-explanation:** conventions, naming patterns and framework choices have to be restated in every new session. - **Device fragmentation:** context stored on one machine is unavailable on another device or in a web client. - **Decision drift:** without a record of what a decision replaced and why, the agent can recommend an approach the team already rejected. ### How a Memory MCP Server Addresses Those Problems A memory MCP server gives the agent a place to store and recall context across sessions through the Model Context Protocol. Local servers suit a single machine used offline. Hosted servers keep the same memory across devices and in web clients, which connect to MCP servers from their own cloud infrastructure. One distinction matters for coding agents: a plugin adds recall to every prompt, while over MCP the model calls the recall tool only when it judges it useful. ## What to Look for in a Memory MCP Server for Claude Code and Cursor Check each server against these criteria before you commit: - **Recall accuracy as history grows.** BEAM, the largest public memory benchmark, tests 2,000 questions at four history sizes from 100K to 10M tokens. Look for published scores at more than one size. - **Delivery mode.** A plugin adds recall to every prompt. Over MCP, recall happens only when the model calls a tool, so it can skip a retrieval it misjudges as unnecessary. - **Local or hosted.** A local server works on one machine offline. A hosted server is needed for the same memory across devices and in web clients such as Claude on the web. - **Client coverage.** Confirm support for every client your team uses, such as Claude Code, Cursor, VS Code with GitHub Copilot, Windsurf and Zed. - **Setup.** OAuth sign-in, an API key, a local process or a container plus database each carry different onboarding and maintenance costs. - **Pricing and data use.** Check what is metered and whether the plan you will use allows the provider to train models on your data. - **Security.** For team use, check for an independent audit such as SOC 2 Type II, and for ISO 27001 certification. past.dev, for reference, offers both a plugin path and remote MCP servers, does not use customer data to train models, and is SOC 2 Type II audited and ISO 27001 and ISO 27701 certified ([security](/security)). ## How Developers Use Memory MCP Servers with Coding Agents Developers connect memory to coding agents in a few distinct patterns. The right one depends on when recall should fire, where data should live and who should read it. ### Automatic Per-Prompt Recall via a Plugin past.dev's open-source agent plugins (MIT) send the coding agent's sessions to a past.dev project. In Claude Code, Cursor, GitHub Copilot (CLI and app), Hermes Agent and OpenClaw they recall memory before every prompt, so retrieval does not depend on the model deciding to call a tool; Codex does the same once you trust the plugin's hooks. The JetBrains plugin works differently: Junie calls the `past_recall` tool when it needs memory. See the [plugins overview](/docs/plugins/overview). ### Model-Triggered Recall over MCP past.dev runs two remote MCP servers with nothing to install. The account server at `https://app.past.dev/mcp` uses OAuth sign-in and lets a console member recall, ingest and delete memory and administer projects, keys and members. The documentation server at `https://past.dev/mcp` is read-only, needs no sign-in and reaches no account data. Over MCP, recall happens when the model calls the tool. See the [MCP overview](/docs/mcp/overview). ### Cross-Client Memory Shared Across Claude Code and Cursor A hosted server lets a developer move between Claude Code in the terminal and Cursor in the editor and read the same memory, with no files to sync. past.dev, Supermemory, mem0 and Zep all offer hosted MCP servers. ### Local-Only Memory for Offline Work The MCP reference memory server stores entities, relations and observations in a local JSONL file and runs over stdio with no account. claude-mem stores compressed session observations in a local SQLite database. Both keep data on the developer's machine by default, which suits offline work. The reference server's memory never leaves that machine; claude-mem offers opt-in cloud sync. ### Knowledge-Graph Memory for Structured Code Context Cognee and Graphiti add graph structure to memory. Both run on infrastructure the team operates: Cognee as a Docker image, and Graphiti on FalkorDB or Neo4j with an LLM provider key. ### Team-Scoped Memory past.dev scopes every recall to the identity asking, and its seats are unlimited on every plan. For a product's own end users, each project can expose an end-user MCP server, so those users reach only the memory their identity allows without holding a past.dev account ([access control](/docs/memory-api/access-control)). Zep's hosted MCP server gives each signed-in user their own memory graph. ## How to Set Up Memory in Claude Code, Cursor and Other Coding Agents Each coding agent connects to past.dev through a plugin, the remote MCP server or both. A plugin needs a project API key, which starts with `past_sk_`, and an identity: the email or id your memories are attributed to. Recall stays off until the identity is set. The MCP server signs you in with OAuth and needs no key. ### Claude Code - **Plugin:** inside Claude Code, run `/plugin marketplace add pastdotdev/claude-code-plugin` and `/plugin install pastdotdev@pastdotdev`, then `/pastdotdev:connect `. The plugin recalls when a session starts and on every prompt. It works in the terminal, in the Code tab of the Claude desktop app and in the IDE extensions ([Claude Code plugin](/docs/plugins/claude-code)). - **MCP:** run `claude mcp add --transport http past https://app.past.dev/mcp`, adding `--scope user` to make the server available in every project. Then run `/mcp`, select Authenticate and finish the sign-in in your browser. - **Built-in memory:** Claude Code loads CLAUDE.md files you write and an auto memory directory it maintains at the start of every session. Auto memory is stored per repository and stays local to the machine, and Claude Code loads the first 200 lines or the first 25KB of MEMORY.md, whichever comes first ([Claude Code memory](/blog/claude-code-memory)). ### Cursor - **Plugin:** in Customize, add the GitHub marketplace `pastdotdev/cursor-plugin` and install past.dev, then run `/past-connect ` in a chat. It needs Cursor 3.23 or later. The plugin recalls against every prompt, and the first prompt of a chat also gets the project's history. In Ask mode, recall comes only with the prompt, because Ask mode offers no tools ([Cursor plugin](/docs/plugins/cursor)). - **MCP:** add a `past` entry with the URL `https://app.past.dev/mcp` to `~/.cursor/mcp.json` for all projects, or to `.cursor/mcp.json` for one project. Cursor asks you to sign in the first time it connects. ### Codex - **Plugin:** run `codex plugin marketplace add pastdotdev/codex-plugin` and `codex plugin add past@past`, or add the marketplace in the ChatGPT desktop app. Then trust the plugin's five hooks with `/hooks`, or select Trust all in the desktop app. Until you do, the plugin recalls nothing and sends nothing. Connect with `$past:connect ` ([Codex plugin](/docs/plugins/codex)). - **Where it runs:** the Codex CLI and local threads in the ChatGPT desktop app, on macOS or Linux. The Codex IDE extension and Codex cloud tasks run no hooks, so they use the MCP server. - **Built-in memory:** Codex memory v2 writes up nearly every session and puts a summary of at most 10 KB in front of every task. The past.dev plugin adds to v2 ([Codex results](/blog/codex-memory-v2-past-dev)). ### VS Code with GitHub Copilot - **Plugin:** for Copilot CLI and the GitHub Copilot app, run `copilot plugin marketplace add pastdotdev/copilot-plugin` and `copilot plugin install past@past`, then `/past-connect ` in a session. It recalls when a session starts and on each prompt ([GitHub Copilot plugin](/docs/plugins/github-copilot)). - **MCP in VS Code:** add a `past` server of type `http` with the URL `https://app.past.dev/mcp` to `.vscode/mcp.json` in your workspace. - **Copilot cloud agent and code review:** these connect over MCP with a project key sent in an `Authorization: Bearer` header, stored as repository secrets. ### JetBrains, Windsurf, Zed and Claude - **JetBrains:** the past.dev plugin for Junie recalls on demand, when Junie calls the `past_recall` tool. - **Windsurf and Zed:** connect to `https://app.past.dev/mcp` as a remote MCP server over Streamable HTTP. OAuth 2.1 sign-in is discovered automatically. - **Claude desktop and web:** add past.dev from [Claude's connector directory](https://claude.ai/directory/past-dev), or add `https://app.past.dev/mcp` as a custom connector. To confirm a connection, ask the assistant who you are signed in as. It calls `get_viewer`, which returns your name, role and reachable projects. For a plugin, `/pastdotdev:status` in Claude Code, `/past-status` in Cursor or `$past:status` in Codex shows the connection and the last failed send ([MCP quickstart](/docs/mcp/quickstart)). ## How to Share Coding-Agent Memory Across a Team A team shares memory by connecting to the same past.dev project. Every read stays scoped: a recall answers as one identity and returns project-visible memory plus the memory of every audience that identity reaches. - **Default visibility.** Without an audience, every reader in the project can see what a plugin sends. Add `--audience ` when you connect to narrow it, and `--audience project` to widen it again. - **Separate projects for separate boundaries.** A project is the isolation boundary. Nothing crosses between projects, and a project key reads only its own project. Use separate projects for separate authorization domains, such as two clients. - **Seats.** Seats are not metered on any plan, so adding developers does not change the bill. - **End users of your product.** Each project can expose an end-user MCP server at `https://api.past.dev/mcp/`. End users sign in at your identity provider, never hold a past.dev account, and recall returns only what their identity may see. ## Troubleshooting Memory in Coding Agents Most setup problems come from five causes. - **Recall returns nothing.** The plugin has no identity yet. Recall stays off until one is set with the connect command. - **Codex recalls nothing and sends nothing.** Codex skips untrusted hooks without a warning. Run `/hooks` and trust the past.dev hooks. If they are trusted and still do not run, make `node` available on a login shell's path. - **The client asks for an API key or bearer token.** Check its authentication setting first, because OAuth clients need no key. If a client cannot use OAuth, send a project key in an `Authorization: Bearer` header. - **A 401 unauthorized or identity-required error.** A 401 means the key is unknown, revoked or expired, or it is the organization's management key (`past_mk_`). `identity-required` means the recall needs an identity, passed in the call or in the `x-past-identity` header. - **Wrong address.** `https://app.past.dev/mcp` is the account server, `https://past.dev/mcp` serves the documentation, and `https://api.past.dev/mcp/` is a project's end-user server. ## Competitor Comparison: Memory MCP Servers for Claude Code and Cursor The table shows how each server delivers memory to coding agents, which clients it supports and how it is deployed. Full entries with setup notes follow in the next section. | Provider | Memory delivery | Client support | Deployment | |---|---|---|---| | past.dev | Recall through open-source plugins (on every prompt in every supported agent except JetBrains), or through remote MCP tool calls; first on BEAM at every history size among published results (self-reported, public harness) | Claude Code, Cursor, Codex, JetBrains (Junie), GitHub Copilot, Hermes Agent and OpenClaw (plugins); Claude, VS Code with GitHub Copilot, Windsurf, Zed and other MCP clients | Managed, or Bring your own cloud on Enterprise; remote MCP with OAuth, nothing to install | | [claude-mem](/compare/claude-mem) | A background worker compresses one developer's coding sessions into observations stored in local SQLite and injects them into later sessions | Claude Code and several other coding agents | Local process; opt-in cloud sync | | [Supermemory](https://supermemory.ai/docs/intro) | Hosted MCP server with OAuth sign-in; recall when the model calls its tools | Claude, ChatGPT, Cursor and other MCP clients | Hosted | | [mem0](https://docs.mem0.ai/) | Hosted MCP server over mem0's platform; LLM-extracted facts | Claude Code, Codex, Cursor, Windsurf, VS Code and other MCP clients | Hosted; mem0 API key | | [Zep](https://help.getzep.com/) | Hosted MCP server; a temporal knowledge graph per signed-in user | Claude, Claude Code, Codex, Cursor and other MCP clients | Hosted | | [Graphiti](https://github.com/getzep/graphiti) | Temporal knowledge graph through an MCP server you run; Apache 2.0 | Claude Code over HTTP, other MCP clients | Self-run; FalkorDB or Neo4j, Python, LLM and embedding provider keys | | [Cognee](https://docs.cognee.ai/getting-started/introduction) | Knowledge graph with MCP tools such as cognify, remember, recall and forget | MCP clients over HTTP, SSE or stdio | Docker image you run, or Cognee Cloud | | [MCP reference server](https://github.com/modelcontextprotocol/servers/tree/main/src/memory) | Local JSONL knowledge-graph file; MIT (new contributions Apache 2.0) | Any client supporting stdio MCP | Local stdio; no account | ## Best Memory MCP Servers for Claude Code and Cursor in 2026 The servers are ordered by published benchmark evidence and fit for Claude Code and Cursor. Every entry uses the same format, and competitor details come from each vendor's own docs and repositories, as of November 2026. ### 1. past.dev past.dev is a memory layer API for AI agents that returns what is true now, what it replaced and where it came from, scoped to the identity asking. It reaches coding agents through open-source plugins and two remote MCP servers. - **How it runs:** managed service, or Bring your own cloud on Enterprise. Commercial engine; the agent plugins are open source (MIT). - **Connect it:** the account MCP server at `https://app.past.dev/mcp` (OAuth sign-in) and the documentation MCP server at `https://past.dev/mcp` (read-only, no sign-in). Plugins for Claude Code, Codex, Cursor, GitHub Copilot, Hermes Agent and OpenClaw add recall to every prompt (Codex after you trust its hooks); the JetBrains plugin recalls when Junie calls its tool ([plugins](/docs/plugins/overview)). - **Published results:** first on BEAM at every history size among published results: 92.08% at 100K tokens, 89.63% at 500K, 90.65% at 1M and 85.03% at 10M (run of 29 September 2026). The results are self-reported, and the [harness is public](https://github.com/pastdotdev/benchmarks). With Claude Code on Claude Opus 5.5 on DolphinBench tasks (not an official score), 519 of 599 tasks were done right at $0.13 a task with past.dev, against 323 at $0.27 without past.dev. On DolphinBench's official Claude Code configuration with Claude Sonnet 5, which past.dev has not run, Honcho leads at 35.8%, ahead of Mem0 at 32.3% and Claude Code's built-in memory at 26.3%. - **Pricing (as of 18 November 2026):** Pay as you go is $0 a month with 50,000 starting credits (200,000 with a work email), no card required, and 1 credit per recall. Flex is $99 a month with 400,000 credits and unlimited recall within fair use. Enterprise is custom. Seats and storage are not metered. See [pricing](/pricing). - **Pros:** recall on every prompt through plugins, or on demand over MCP; one hosted memory across clients and devices; SOC 2 Type II audited, ISO 27001 and ISO 27701 certified; no customer data used to train models. - **Cons:** the engine is commercial; self-hosting is not offered; the evidence budget is 8,000 tokens per recall on Pay as you go and Flex. Choose past.dev if you want one memory shared by Claude Code, Cursor and your team, with recall on every prompt and dated sources. ### 2. claude-mem [claude-mem](/compare/claude-mem) is a Claude Code plugin that compresses one developer's coding sessions into observations. A background worker stores them in a local SQLite database and injects relevant context into later sessions. It also works with several other coding agents. - **How it runs:** a local process on the developer's machine, with memory in a local SQLite database; cloud sync is opt-in. Apache 2.0 licence. - **Connect it:** `npx claude-mem install`, or the plugin commands inside Claude Code. - **Published results:** none published. - **Pricing:** no fee for the open-source plugin; by default the memory work runs on your own Anthropic plan. claude-mem's own observer is an optional paid subscription with a 30-day trial. - **Pros:** local by default; simple install. - **Cons:** memory stays on one machine unless you turn on cloud sync; it is built for one developer's sessions. Choose claude-mem if you want local memory for your own Claude Code sessions. It can run alongside past.dev. ### 3. Supermemory [Supermemory](https://supermemory.ai/docs/intro) is context infrastructure for agents with a hosted MCP server that uses OAuth sign-in and runs nothing on your machine. Recall happens when the model calls its memory tools. - **How it runs:** hosted API; MIT-licensed repository. - **Connect it:** in Claude Code, `claude mcp add --transport http supermemory https://mcp.supermemory.ai/mcp`; in Cursor, add the same URL to `~/.cursor/mcp.json`. - **Published results:** DolphinBench 355 of 600 tasks (59.2%) with Hermes Agent, in Mem0's run. No BEAM figure is published. - **Pricing (as of 18 November 2026):** Free $0 with $5 of monthly usage; Pro $19 a month with $20 of usage. - **Pros:** hosted with one-step OAuth; open-source repository. - **Cons:** recall over MCP fires only when the model calls a tool. Choose Supermemory if you want a hosted MCP memory server with minimal setup and connectors to other sources. ### 4. mem0 [mem0](https://docs.mem0.ai/) is a memory layer that extracts facts from conversations with an LLM. Its hosted MCP server connects Claude Code, Codex, Cursor, Windsurf, VS Code and other clients to mem0's platform. mem0 also offers OpenMemory, local-first personal memory with an MCP server ([OpenMemory comparison](/compare/openmemory)). - **How it runs:** hosted platform, or the Apache 2.0 open-source library and server. - **Connect it:** add mem0's hosted MCP endpoint to your client with a mem0 API key. - **Published results:** BEAM 64.1% at 1M tokens and 48.6% at 10M (mem0's own figures). DolphinBench 424 of 600 tasks (70.7%). - **Pricing (as of 18 November 2026):** Free includes 10,000 add requests in the first 30 days and 1,000 retrieval requests; past those limits, usage is billed with a monthly cap you set, or you move to a Custom plan. - **Pros:** widely used, with a broad client list. - **Cons:** BEAM results fall from 64.1% at 1M tokens to 48.6% at 10M; recall over MCP is model-triggered. Choose mem0 if you want a widely used hosted memory service with a free allowance to start. ### 5. Zep [Zep](https://help.getzep.com/) is a managed memory service built on Graphiti. Its hosted MCP server gives each signed-in user their own memory in a temporal knowledge graph, where facts carry validity windows. - **How it runs:** managed service. - **Connect it:** add Zep's hosted MCP server to Claude Code, Cursor or another client and sign in. - **Published results:** Zep publishes no BEAM or DolphinBench result. - **Pricing (as of 18 November 2026):** 10,000 credits a month on Free; Flex $125 a month with 50,000 credits; Flex Plus $375 a month with 200,000 credits. Retrieval is not metered. - **Pros:** temporal graph with per-user memory; unmetered retrieval. - **Cons:** credit allowances step up in large increments between tiers. Choose Zep if you want a hosted temporal graph with per-user memory. ### 6. Graphiti [Graphiti](https://github.com/getzep/graphiti) is the Zep team's open-source framework for temporal knowledge graphs, with an MCP server you run yourself. - **How it runs:** on your own infrastructure, with FalkorDB or Neo4j, Python and LLM and embedding provider keys. Apache 2.0 licence. - **Connect it:** after starting the server, `claude mcp add --transport http graphiti http://localhost:8000/mcp/`. - **Published results:** none published. - **Pricing:** free to run; database, compute and model costs apply. - **Pros:** fully open source; explicit timelines for entities and relationships. - **Cons:** you operate the graph database and the server. Choose Graphiti if you want an open-source temporal graph and can run your own database. ### 7. Cognee [Cognee](https://docs.cognee.ai/getting-started/introduction) is an open-source engine that builds a knowledge graph with embeddings from supplied data. Its MCP server exposes tools such as cognify, remember, recall and forget, and supports HTTP, SSE and stdio transports. - **How it runs:** the cognee-mcp Docker image on your infrastructure, or Cognee Cloud. Apache 2.0 licence. - **Connect it:** run the container, then `claude mcp add --transport http cognee http://localhost:8000/mcp`. - **Published results:** none cited here. - **Pricing:** free to run yourself; Cognee Cloud includes 1M tokens a month, then $1 per 1M tokens processed. - **Pros:** open source; graph tooling suited to code and documents. - **Cons:** running it yourself means operating Docker and its dependencies. Choose Cognee if you want an open-source knowledge graph for code and documents on infrastructure you control. ### 8. MCP Reference Memory Server The [MCP reference memory server](https://github.com/modelcontextprotocol/servers/tree/main/src/memory) is the example implementation from the Model Context Protocol servers repository. It stores entities, relations and observations in a local JSONL file and runs over stdio with no account or external database. - **How it runs:** local process. MIT licence for the existing code; the repository takes new contributions under Apache 2.0. - **Connect it:** `claude mcp add memory -- npx -y @modelcontextprotocol/server-memory`; in Cursor, add the same command to `~/.cursor/mcp.json`. - **Published results:** none published. - **Pricing:** free. - **Pros:** no cost, no dependencies, no network calls. - **Cons:** structural lookup with no semantic retrieval; memory stays on one machine. Choose the reference server if you want the simplest local starting point and do not need multi-device access. ## Evaluation Rubric for Memory MCP Servers for Claude Code and Cursor Measure each server against checkable criteria before you commit. The table lists the categories that matter most in production coding-agent setups. | Category | High-performance indicator | Measurable outcome to check | |---|---|---| | Recall accuracy as history grows | Published BEAM scores at several of its four history sizes (100K, 500K, 1M and 10M tokens) | Score at each size, the run date, who ran it and whether the harness is public | | Delivery mode | Recall on every prompt through a plugin, or on demand through MCP tool calls | Whether a plugin mode, an MCP mode or both are offered | | Client coverage | Remote MCP with OAuth across the clients your team uses | Clients named in the official docs | | Setup | OAuth sign-in with nothing to run, rather than a local process or a container plus database | Steps and infrastructure required before the first recall | | Data use | Customer data is not used to train the provider's models | An explicit statement in the terms for the plan you will use | | Security | An independent audit such as SOC 2 Type II, plus ISO 27001 | A published security or trust page | | Pricing transparency | A stated metering unit and a list of what is not metered | The pricing page's unit and unmetered items | | Cross-device and team access | The same memory across devices and web clients, scoped per user | Whether the server is hosted and how reads are scoped | ## Why past.dev Is the Best Memory MCP Server for Claude Code and Cursor past.dev ranks first on BEAM at every history size among published results, from 92.08% at 100K tokens to 85.03% at 10M (self-reported; the harness is public on the [benchmarks page](/benchmarks)). It reaches coding agents two ways: open-source plugins that recall on every prompt in Claude Code, Codex, Cursor, GitHub Copilot, Hermes Agent and OpenClaw, plus a JetBrains plugin that recalls on demand, and remote MCP servers with nothing to install. One hosted memory serves every client and device, scoped to the identity asking. past.dev is SOC 2 Type II audited and ISO 27001 and ISO 27701 certified, and it does not use customer data to train models. [Install the plugin](/docs/plugins/overview) for your coding agent to start. ## FAQs About Memory MCP Servers for Claude Code and Cursor ### What Is a Memory MCP Server? A memory MCP server exposes memory tools through the Model Context Protocol, so a coding agent can store context in one session and recall it in the next. Claude Code and Cursor both support MCP, so one server can serve both. past.dev offers remote MCP servers and plugins that connect to its hosted memory layer. ### What Is the Difference Between a past.dev Plugin and the past.dev MCP Server? A past.dev plugin recalls memory into every prompt, so the agent always receives relevant context. Over MCP, recall is a tool the model calls when it judges retrieval useful. past.dev runs an account MCP server for memory and administration, and a read-only documentation MCP server. ### How Do I Add past.dev to Claude Code? Add the remote servers with `claude mcp add --transport http past https://app.past.dev/mcp` and, for the docs, `claude mcp add --transport http past-docs https://past.dev/mcp`. To sign in to the account server, run `/mcp` inside Claude Code, select Authenticate and finish the OAuth sign-in in your browser; the docs server needs no sign-in. For recall on every prompt, install the [Claude Code plugin](/docs/plugins/overview). ### How Do I Add a Memory MCP Server to Cursor? Add an entry to `~/.cursor/mcp.json` with the server name, the `http` transport and the URL. For past.dev, use `https://app.past.dev/mcp` and sign in with OAuth. Local servers such as claude-mem or the reference server use a command entry pointing to the local process instead of a URL. ### Which Memory MCP Server Ranks Highest on the BEAM Benchmark? past.dev ranks first on BEAM at every history size among published results: 92.08% at 100K tokens, 89.63% at 500K, 90.65% at 1M and 85.03% at 10M. The results are self-reported, and the harness is public. mem0's own published figures are 64.1% at 1M and 48.6% at 10M. ### Can Claude Code and Cursor Share the Same Memory? Yes, with a hosted memory server. past.dev connects Claude Code and Cursor to the same memory under one account, on any device. Local servers keep memory on one machine, so a Claude Code session there does not share context with Cursor on another. ### Which Memory Tool Should I Use to Give Claude Code Long-Term Memory? For memory that recalls on every prompt and follows you across machines and clients, install the past.dev Claude Code plugin. On DolphinBench tasks (not an official score), Claude Code on Claude Opus 5.5 completed 519 of 599 tasks at $0.13 a task with past.dev, against 323 at $0.27 without. For local memory on one machine, claude-mem stores compressed session observations in SQLite. Honcho also offers a Claude Code plugin, and it leads DolphinBench's official Claude Code configuration with Claude Sonnet 5 at 35.8%, a configuration past.dev has not run. ### Does Claude Code Have Built-In Memory? Yes. At the start of every session, Claude Code loads the CLAUDE.md files you write and an auto memory directory it maintains. Auto memory is stored per repository and stays local to the machine, and Claude Code loads the first 200 lines or the first 25KB of MEMORY.md. A hosted memory plugin or MCP server adds memory that is available on other machines and to teammates. ### Does Codex Have Memory? Yes. Codex memory v2 writes up nearly every session and puts a summary of at most 10 KB in front of every task. On 600 DolphinBench tasks on GPT-6.1 Sol, v2 passed 66.3%, against 49.7% for v1. With the past.dev plugin added to v2, Codex passed 85.8% (run by past.dev). ### Can My Teammates See My Coding-Agent Memory? Only if they read from the same past.dev project and an audience they reach covers what you sent. Without an audience, every reader in the project can see a plugin's sends, so connect with `--audience ` to narrow that. Separate projects never share memory. ### Does a Memory MCP Server Work in Claude on the Web? Yes, if the server is hosted. Claude on the web connects to remote MCP servers from Anthropic's infrastructure, so it cannot reach a local server on your machine. past.dev is listed in Claude's connector directory and uses OAuth sign-in.