A lead engineer whose people split work across Claude Code, Cursor, Codex and JetBrains needs a place where one decision stays visible when the next person opens a different tool. This article describes what each vendor documents; it does not test products.
TL;DR
- Three kinds of product answer the question. Shared memory stores the team writes to: Basic Memory, Cognee, Mem0, MemoryLake, MemoryPlugin, Mnemoverse (the author's company), Supermemory and Zep in the vendor's cloud, or TencentDB Agent Memory on the team's own servers. Code and company context served to agents over MCP: Augment Context Engine, Glean, Sourcegraph and Unblocked. Team rules and skills rendered into each tool's files: Packmind, Augment workspace rules and BuildBetter ZeroShot.
- Every tool still gets its own connection. In every option checked, each coding tool needs its own connector or its own file, with its own sign-in where the vendor uses one. Unblocked, Glean and Packmind automate parts of that setup.
- Check who sees a teammate's contribution before comparing prices. Cognee, Letta, Supermemory and Zep document four different defaults, and for Supermemory and Zep, narrowing them takes a higher plan.
- One machine is a different problem from a team. Tencent's MemHub, published on 16 October 2026, runs one runtime per installation; Tencent's team product is a different one, TencentDB Agent Memory.
The full comparison table, with plans, prices and read dates, is in the library version.
Three kinds of cross-IDE context platform
A shared memory store is a persistent knowledge base that people and their agents write to and each connected tool reads back. Code and company context is an index of code, pull requests, documents, tickets and chat that agents search over MCP (Model Context Protocol, through which a coding tool reaches an outside server's tools). Team rules and skills are standards and skills kept in one place and written into the instruction file each assistant already reads.
Start with the missing artifact: a decision from a coding session points to memory, an answer already in a repository or ticket to context search, a coding convention to rules. The connection is a separate concern: each tool still receives its own MCP entry, plugin, extension, proxy setting or instruction file. A tool not named on the pages read is not thereby unsupported.
Tencent's MemHub: one installation versus a team
On 16 October 2026 Tencent Hunyuan published the open repository Hy-Memory, whose first commit is titled "Initial open source release". Tencent's MemHub, built on it, is "a local-first desktop and CLI product that unifies conversation history and coding traces from user-authorized AI applications", and its architecture page says "There is one managed Runtime and one MemHub data home per installation." Its support table dated 14 October marks eight of ten tools as open, Claude Code, Cursor and Codex among them.
Releases v1.0.0 and v1.0.2 of 16 October and v1.0.4 of 18 October carry an Apple silicon Mac app and Linux runtimes, none an Intel Mac build, and on 18 October the README announced memhub-cli on npm. Windows is not supported in any of the three; a maintainer said support would come, with no date. The connection prompt tells the agent that web chats, remote containers and another machine cannot use these local paths, and the repository and site describe no hosted MemHub service as of 18 October.
Shared memory stores for engineering teams
Each tool reaches the store through its own connector (the mechanism); what differs for a team is the scope behind the connection.
Basic Memory connects over remote MCP or local stdio. In Basic Memory Teams, "Anything a teammate writes is immediately available to everyone else and to their agents", under five roles; the local version has "No team features; runs entirely on your computer".
Cognee Cloud runs a hosted MCP server with OAuth, and "recall without a dataset searches everything you can access, so memories written by one client are visible to the others." Self-hosted, the open cognee-mcp shares one graph only in API Mode.
Letta keeps memory inside an agent: through its ACP adapter, "the agent and its memory stay on the Letta Platform while tools run against the working tree on your machine", in Zed, JetBrains IDEs or Obsidian. Its hosted MCP server is for delegating work to Letta agents, and per its permissions page, "Agents are private to their creator by default."
Mem0's MCP page names seven example clients for its hosted server, Claude Code, Codex and Cursor among them, and in its Claude Code integration "Everyone on the repo reads and writes the same pool."
MemoryLake connects over MCP, and "you select the workspace and boundary on the OAuth consent screen when a client connects" (concepts). Its Claude Code page says "Paid tiers only for team collaboration."
MemoryPlugin reaches ChatGPT, Claude and Gemini through a browser extension and Claude Code or Cursor through a hosted MCP server; its Pro card reads "Shared buckets for your team".
Mnemoverse documents one memory per account: "Any key or sign-in from the same Mnemoverse account reaches the same memory." For a team, "Each engineer keeps their own API key; the team shares knowledge through a room (Beta)". The limit to weigh: rooms are in beta.
Supermemory's coding plugins for Cursor, Claude Code, Muse Code, Codex and OpenCode share one repository tag, "so agents working on the same repo read and write the same memory". Team invitations start on Pro, an invited Member gets all containers by default, and limiting a teammate to selected container tags needs Scale or Enterprise.
TencentDB Agent Memory is "a team-level memory hub for AI Agents" from Tencent Cloud that, per its repository, a team deploys itself. An agent points its model base URL at the team's proxy: "No plugin, hook, or MCP server is required." Every LLM request then passes through that proxy. New chat memory and skills are private by default, and the visibility settings are private, team, restricted and agent.
Zep's Context MCP Server is one address for every client, and per its graph authorization page the default is broad: "Every authenticated MCP user can list and search every shared Context Graph in the token-bound project." Moving connections to UserGroup ABAC on Zep managed cloud requires Enterprise and the ABAC entitlement.
Who sees what on a shared memory store
For a team this matters more than the connector: what can a newly connected engineer search before an administrator changes anything? On Zep, every shared graph in the project; on Supermemory, every container; on Cognee Cloud, all tenant-wide datasets for invited members, with access control per dataset. Letta agents and new TencentDB Agent Memory chat memory and skills are private by default; on Mnemoverse, each engineer has their own account plus what the team shares through a room (beta). Pilot-test these defaults before sensitive material goes in.
Code and company context over MCP
These answer a different question: what existing evidence should the agent read?
Augment Context Engine serves code context over MCP two ways: "Run the Auggie CLI locally as an MCP server (stdio)", or a hosted engine that indexes "Selected repos' default branches (chosen during GitHub App install)". Its docs carry 12 quickstarts, Claude Code, Codex and Cursor among them.
Glean brings permission-aware enterprise context to MCP hosts, "so users can only search, read, and act on content they can already access in Glean", with a plug-in for Cursor, Claude Code and Codex; separately, administrators can push MCP configuration to supported hosts through MDM.
Sourcegraph's MCP server is "Supported on Enterprise plans.", and "Users can only read data from repositories they are already allowed to access in Sourcegraph."
Unblocked indexes code, pull requests, Slack, Confluence, Jira, Linear and more. "The install script supports macOS and Linux." (MCP setup), and its Shared Memories keep corrections from shared conversations, managed by admins.
Team rules and skills across coding tools
Here the team shares instructions rather than a store, written once and rendered into the file each assistant reads; a team can use both.
Augment workspace rules live in .augment/rules, "for project-specific guidelines that should be shared with your team via version control." User guidelines are a different scope: "Guidelines defined in VSCode will not propagate to JetBrains IDEs and vice versa."
BuildBetter ZeroShot "turns your team's sessions and PR reviews into reusable skills", and client lists differ by plan card: the first says "Works with Claude & Codex", the Pro card "Works with Claude, Codex, Cursor & Gemini".
Packmind keeps its sources in .packmind/ as "the single source of truth" and renders them into each assistant's own files, 10 targets in the vendor's table, from Claude Code, Cursor and GitHub Copilot to AGENTS.md and Kiro.
Local-first context on one developer's machine
Local-first is a deployment property, not a substitute for inspecting the sharing controls.
ByteRover V4 "connects agents through the published ByteRover skill.", and in the vendor's words, "Your memory stays local first, with a cloud option to selectively share it with other agents and people."
claude-mem runs a local worker with SQLite and Chroma, by default "limits startup context to the current harness", and offers TeamBrain, "An optional shared workspace beside My Work", where "Each person chooses what to share"; its v13.40.0 release of 16 October added joining a team from the terminal.
Pieces keeps memory on the device, "All captured context is processed and stored entirely on your device", with organization policy on Enterprise. Tencent's MemHub belongs here too.
How to choose a cross-IDE context platform for a team
What should persist, and who writes it? Basic Memory takes what teammates write, Sourcegraph serves existing code, Packmind distributes maintained guidelines, and ZeroShot derives skills from sessions.
Where may the data live? On your own servers: TencentDB Agent Memory (a model-request proxy), cognee-mcp in API Mode (a shared graph), Mem0 Open Source (a REST API with a dashboard) and self-hosted Packmind (rules).
Who sees a new contribution, and from which plan? SSO is on Basic Memory Business, SSO and SCIM on Unblocked Enterprise, UserGroup ABAC on Zep Enterprise, per-tag limits on Supermemory Scale or Enterprise, and MCP itself on Sourcegraph Enterprise plans.
Must browser chats use the same context? Cognee names Claude on the web and ChatGPT, Supermemory documents ChatGPT developer mode, Zep has an administrator quickstart for Claude and ChatGPT, and MemoryPlugin uses a browser extension; Tencent's MemHub prompt says web chats cannot use its local paths.
Does anyone work on Windows? Windows is not supported in Tencent's MemHub v1.0.0, v1.0.2 or v1.0.4, and Unblocked's install script covers macOS and Linux, with a guide for each app on Windows.
For a pilot, follow one contribution from authoring to retrieval in another teammate's tool, checking attribution, access scope and billing ownership.
Sources and method. The vendor pages linked above were read on 18 October 2026, every quotation with two instruments (a browser-like request and a second reader; for GitHub, the API and the raw file or page), with a made-up address on each host answering 404 or a not-found page as a control. The text describes these products as their public pages stood that day; releases, plans and tool lists change, so check the vendor's page before deciding. Mnemoverse, too, is described only from its public documentation. This is a reading of documentation, not a hands-on test of any product.
Edward Izgorodin · Mnemoverse · 2026-10-18
Top comments (0)