[Feature Request] Unified user state across Cowork and Claude chat — shared memory, files, skills, connectors
Problem
Claude Cowork (Desktop research preview) and Claude chat (web + iOS + Android) operate as siloed surfaces despite serving the same authenticated user:
- Cowork can read/write a scoped folder on my Mac, run skills, use plugins, and execute multi-step file-centric work — but only at my desk, with no mobile equivalent.
- Chat (especially the iOS app) is where I actually spend time when I'm away from my desk. It has its own memory, its own connectors, its own thread history.
When I shift surfaces during a workflow, context resets. Memory doesn't follow. Files I worked on in Cowork aren't visible in chat on the go. Connector state and authentications don't carry. Skills configured for one don't appear in the other (related: #20697).
Use case
Typical day: morning at desk → Cowork on a project folder → commute → continue thinking in iOS chat → evening back at desk → Cowork to execute. Today this requires manual context-passing. I use Claude Code Dispatch from iOS to route specific tasks to a Claude Code session on my Mac, which works well, but it doesn't solve the underlying memory and file split — it just routes a single task across.
I've worked around this by making Obsidian Sync my canonical file backend (Claude Code reads it directly on Mac, Obsidian Sync handles mobile access), but that only partially closes the gap and doesn't address memory or connector state.
Suggested approach
Treat Cowork and chat as two surfaces over a single unified user state:
- Shared memory — claude.ai memory, Cowork auto-memory, and Claude Code CLAUDE.md should converge into one coherent layer the user can read/write from any surface.
- Shared file workspace — when Cowork has a folder scoped, that folder (or selected artifacts) should be reachable from chat as a viewable/searchable resource. From iOS I should be able to ask "what's the latest on the X analysis I was working on in Cowork?" and get an answer grounded in Cowork's actual files.
- Shared skills and connectors — already requested for skills in #20697; extend to connectors and authentication state.
- Surface handoff affordances — explicit "continue this in Cowork" / "continue this in chat" actions, analogous to Claude Code's
/desktopand--teleportfor the coding case.
Related
- #15542 — Claude Code ↔ Claude App chat history
- #2511 — Claude Code ↔ Claude Projects
- #20697 — Sync Skills between Desktop and CLI
- #14228 — Claude Code access to claude.ai memory
This issue is specifically about the Cowork ↔ chat axis (file-centric desktop work ↔ on-the-go mobile chat), which the issues above don't quite cover.
Showing cached comments. Read the full discussion on GitHub ↗
8 Comments
Found 2 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
Downvoted auto close since each alleged dupe report makes the case somewhat differently and offers alternative perspectives. Consider their aggregate, merged.
I am _guessing_ one of the barriers to the obvious sharing is down to different quotas and plans. That Anthropic are kind of relying on the accidental profusion of pricing levers to triage expensive compute.
But if so: it being so hard to move context is the problem with your current ecosystem, folk. If you can't afford the ideal amalgamation, at least write a few more conversion tools that can reach into the innards of systems. (I will cough again that exposing more MCP affordances to share within a single account is one possible lightweight approach.)
The paper cuts when we DIY this stuff make our fingers bleed.
❤️
+1, particularly on axis 1 (shared memory). The split between claude.ai memory and Claude Code's
CLAUDE.mdis the daily pain point — same authenticated user, two separately-curated knowledge layers that drift on every session.I'm working around this user-side by mirroring my claude.ai memory entries verbatim into a Notion page, which Claude Code can then query through the Notion MCP. It functions, but the round-trip (memory edit → Notion sync → CC fetch) is fragile, and every memory mutation needs manual hook-firing to stay coherent. I filed a related proposal with Anthropic privacy@ in April for a Memory Export API as a minimum primitive — even a read-only export over OAuth would let user-side syncs converge without each of us reinventing the mirror.
Axes 2–4 are real but mostly downstream of axis 1: if memory unifies, much of CLAUDE.md / skills could derive from it, and connector state is largely an auth-scope question. Surface handoff (4) is probably the most user-visible win and the easiest first product surface — a
continue this in Coworkaction from chat (or vice versa) would close most of the daily friction even before the deeper unification ships.Love this, cowork artifacts should be visible on mobile read only, even without linking devices which is risky. Really surprised it's not supported.
I run a small remedial construction/waterproofing business in Perth, and I'm also building a side project (a training app) using Claude Code. I use claude.ai chat, Claude Code, and Cowork all in the same week, sometimes the same day, on the same work.
Right now I keep a CLAUDE.md file as my own workaround — basically a manual "memory" file I update and re-paste so each tool knows what's going on. It works, but it's on me to keep it in sync and remember to feed it to whichever tool I'm in. If I forget, I'm re-explaining myself from scratch.
I'm not a developer, just someone using these tools daily to run a business. From where I sit it seems like the fix doesn't need to be complicated — even something simple like: if a file like CLAUDE.md sits in a connected Drive/project folder, let any Claude surface (chat, Code, Cowork) auto-check and load it at the start of a session, instead of needing me to ask every time. Doesn't have to be deep memory-merging, just "read the same file automatically."
+1 for this. Even a basic version would save a lot of repeated typing for people stitching these tools together day to day.
Need this - kind of crazy this hasn't been done yet by the anthropic team
Strongly behind this. The one surface I'd explicitly fold into the unified state is voice. Right now voice is the most fragmented of all. My Cowork and Claude Code CLI sessions carry my whole second brain, my vault, my memory, my accumulated context, but the moment I open a voice conversation, in Cowork or on mobile, all of that vanishes and I'm talking to a stranger. If context handoff becomes seamless across surfaces, voice has to be a first-class citizen in that set, not an afterthought. Being able to say, continue this in voice, and have my full context come with me, would be the single biggest unlock. I don't want to be the glue anymore, especially not when my hands are busy and voice is the only way I can work.
You basically wrote the spec here — memory that follows the user across surfaces instead of living in per-app silos. Rather than wait for Anthropic to ship it, I built the boring version that works today: Silicon Cortex, a local plain-Markdown context layer that every tool you use reads, so context survives surface switches because it's files you own — nothing resets when you change apps. Just launched a $59 founding edition — 50 seats, priced low in exchange for feedback from people who've hit this exact wall. Want the link?