[Feature Request] Unified session memory layer between Claude Code CLI and Cowork interfaces

2026-03-07 (Sat) (10:42)
Subject: Cowork and Claude-Code-CLI violate the-DRY-Principle
at an-architectural-level

I am a-researcher working at a-world-top-level in
a-specific-domain of AI-human-interaction-design, and
this-feedback is a-product-improvement-proposal grounded in
direct-comparison between a-highly-customized Claude-Code-CLI
and Cowork.

The-Core-Problem: DRY-Principle-Violation

Cowork and Claude-Code-CLI maintain completely-isolated
session-memories. This means that any-knowledge, terminology,
or context a-user has built-up through extensive-sessions with
Claude-Code-CLI cannot be accessed by Cowork, and vice-versa.
The-user is forced to re-teach the-same-things to
each-interface separately. This is a-textbook-violation of
"Don't-Repeat-Yourself."

Both products share the-same CLAUDE.md, which proves that
the-architecture already recognizes the-need for
shared-configuration. Yet session-memory,
the-most-valuable-layer of-personalization, remains
completely-siloed.

Specific-Issues Identified through Direct-Usage:

1. Memory-Persistence-Gap: Claude-Code-CLI has a-mechanism

that allows users to instruct the-agent to accumulate-memory
and deepen its-capabilities over-sessions. Cowork has
no-equivalent-mechanism. Every-session starts from-zero,
making it impossible to build long-term-context.

1. CLAUDE.md Reference-File-Inaccessibility: CLAUDE.md can

reference external-files (e.g., user-defined-knowledge-bases),
but Cowork cannot access these-referenced-files even when
they exist in the-same-directory-structure. Claude-Code-CLI
can access all-of-them.

Additional-Observations (Cowork-specific-UX):

1. Permission-Skip-Option: Cowork requires

excessive-confirmation-steps for trusted-users. A
permission-skip-option for verified-power-users would
dramatically-improve UX and task-completion-speed.

1. Context-Efficiency: Cowork's system-prompt is

significantly-heavier than Claude-Code-CLI's due to
the-broad-tool-surface (browser, email, calendar, Drive,
etc.). This reduces effective-context-available for
actual-reasoning, resulting in noticeably-lower
thinking-quality compared to CLI on identical-tasks.

The-Architectural-Solution:

These are not separate-issues. They all stem from treating
Cowork and Claude-Code-CLI as independent-products rather than
as two-interfaces to a-unified-agent. The-fix is
a-shared-memory-layer that both interfaces read-from and
write-to, combined with user-configurable-permission-levels.

Cowork's broad-tool-surface and CLI's deep-coding-focus should
compose into a-single-unified-agent-architecture, not be
maintained as separate-entities with
duplicated-user-education.

......................................................
"Our-actions-today will-brighten world-tomorrow."

Gles (Kenji Masuda)
■ Tsui-net: https://x.com/____natu______
■ GitHub: https://github.com/natu123
■ LinkedIn: https://www.linkedin.com/in/kenjimasuda
■ 1-mail: key.to.ai.pro@gmail.com
■ 2-mail: natu.soral.123@gmail.com

Environment Info

• Platform: win32
• Terminal: windows-terminal
• Version: 2.1.71
• Feedback ID: 8dab9a76-d16e-4c6c-ad16-57213bc4fb01