Feedback: no shared state across Claude surfaces/sessions; long conversations also degrade
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Area: Cross-session / cross-surface continuity; long-context durability
Summary
Claude has no shared point of contact or shared state across its surfaces — Claude Code (CLI), the browser extension, claude.ai chats, and separate projects. Each session is isolated and cannot see the others. For multi-step tasks that unfold over time, this forces the user to be the only memory and the only coordinator — exactly when help is most needed.
What happens in practice
- A task gets split across multiple sessions/surfaces that can't see each other's work.
- Two sessions can edit the same files/workspace at once, producing conflicting actions and duplicate outputs, with no awareness another session is active.
- A fresh session has no idea what prior sessions already produced, so it rediscovers or contradicts earlier work.
Related: single long conversations also degrade
The context window is finite, so long/involved chats get compacted and earlier details quietly drop out (decisions, file paths, naming, caveats). The user experiences this as "forgetting" or losing the thread — and it hits hardest on the long, multi-step tasks that span the most tokens.
Requested improvements
- Shared, persistent project memory any surface can read/write — one source of truth, not silos.
- Awareness of other active sessions on the same workspace, with a soft lock to prevent simultaneous conflicting edits.
- An auto-maintained project state/index so a new session sees what exists and what's been done.
- Continuity across Code / extension / chats / projects referencing the same context.
- Within a session: checkpoint important state to durable storage before compaction, warn as context fills, and keep a decisions/artifacts ledger that survives summarization.
Severity: High for anyone doing real, multi-piece work. Per-session capability is strong; the gap is coordination across sessions/surfaces and durability within a long session.
Proposed Solution
See requested improvements in Problem Statement above:
- Shared, persistent project memory any surface can read/write — one source of truth, not silos.
- 2. Awareness of other active sessions on the same workspace, with a soft lock to prevent simultaneous conflicting edits.
- 3. An auto-maintained project state/index so a new session sees what exists and what's been done.
- 4. Continuity across Code / extension / chats / projects referencing the same context.
- 5. Within a session: checkpoint important state to durable storage before compaction, warn as context fills, and keep a decisions/artifacts ledger that survives summarization.
Alternative Solutions
Currently the only workarounds are manual and burdensome: copy-pasting relevant context into each new session by hand, maintaining a CLAUDE.md file as a makeshift shared memory, or keeping external notes/docs to track decisions and file state across sessions. None of these scale for real multi-step work.
Priority
High - Significant impact on productivity
Feature Category
Other
Use Case Example
- I start a Claude Code session to build a new feature — it produces files, makes decisions about naming and structure, and gets partway through.
- 2. The session hits context limits and I have to start a new one.
- 3. The new session has no idea what the previous one did — it rediscovers the codebase from scratch, may contradict earlier naming decisions, or duplicates work already done.
- 4. Separately, I open a claude.ai chat to think through the architecture. That chat cannot see what Claude Code already built.
- 5. If I accidentally run two Claude Code sessions on the same workspace, they edit the same files simultaneously with no awareness of each other, producing conflicts.
- 6. At every step, I am the only coordinator — manually summarizing prior work into each new session just to keep things coherent.
Additional Context
_No response_
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗