Post
120
Quick observation from working with Claude Cowork on multi-session projects.
**The constraint:** Cowork runs in a sandboxed VM that resets between sessions. No state persistence, no conversation carryover, no cross-session memory. Each session starts with zero context about prior work.
Within a single session, context compaction occurs as conversations approach the token limit — the system summarizes earlier exchanges, trading detail for capacity. Decisions made early in a session degrade to single-sentence summaries as the conversation grows.
**The cost:** For task-based work (summarize this, debug that), this is irrelevant. For cumulative work — projects that build over days or weeks — it creates significant overhead. Users spend meaningful time each session re-establishing context that the AI previously held.
**Common workarounds observed:**
- Manual context injection (copy-paste at session start)
- Single mega-sessions (avoids reset, but context degrades)
- External state documents (maintained alongside the AI)
Each trades one form of overhead for another.
**The architectural question:** Claude's chat interface (claude.ai) now has persistent memory — free for all users since March 2026. But the desktop environment still operates on a session-reset model. The gap between chat-based memory and desktop-based amnesia is growing.
This seems like a general problem for desktop AI tools, not just Cowork. How are others approaching session continuity in local AI environments?
**The constraint:** Cowork runs in a sandboxed VM that resets between sessions. No state persistence, no conversation carryover, no cross-session memory. Each session starts with zero context about prior work.
https://craftframework.ai/the-hidden-cost-of-session-amnesia-why-context-matters-more-than-you-think/Within a single session, context compaction occurs as conversations approach the token limit — the system summarizes earlier exchanges, trading detail for capacity. Decisions made early in a session degrade to single-sentence summaries as the conversation grows.
**The cost:** For task-based work (summarize this, debug that), this is irrelevant. For cumulative work — projects that build over days or weeks — it creates significant overhead. Users spend meaningful time each session re-establishing context that the AI previously held.
**Common workarounds observed:**
- Manual context injection (copy-paste at session start)
- Single mega-sessions (avoids reset, but context degrades)
- External state documents (maintained alongside the AI)
Each trades one form of overhead for another.
**The architectural question:** Claude's chat interface (claude.ai) now has persistent memory — free for all users since March 2026. But the desktop environment still operates on a session-reset model. The gap between chat-based memory and desktop-based amnesia is growing.
This seems like a general problem for desktop AI tools, not just Cowork. How are others approaching session continuity in local AI environments?
