Quality regression: no cross-session context retention, repeated failures on same task

Summary

Claude Code systematically loses operational context between sessions on the same ongoing task, causing it to repeat the same mistakes, reopen the same problem multiple times, and fail to capitalize on previous errors within the same working day.

Environment

• Claude Code CLI (latest)
• Multi-machine homelab + professional sysadmin environment
• User: Linux sysadmin, 19 years experience

Pattern observed (June–July 2026)

XFCE panel widget task — 9 sessions, same failures repeated

Between 2026-06-30 16:24 and 19:51, 9 separate Claude sessions were opened for the
same task (installing a genmon widget on XFCE4 panel, Ubuntu 24.04).

In each new session, Claude:

• Started from scratch with no memory of what had been tried
• Applied the same incorrect commands that had already failed
• Did not acknowledge or reference the previous session's outcome
• Made the same configuration mistakes (pkill loops, xfconf/rc conflicts)

GNU Screen terminal display corruption — 5 sessions, fix announced but not verified

Between 2026-07-01 12:26 and 13:54, 5 sessions were opened for terminal display
corruption in GNU Screen (ANSI color codes rendering as artifacts).

• Claude applied a fix (.screenrc, TERM=xterm-256color) and announced the problem solved
• In each following session, the same problem was reported again
• Claude never verified that the fix actually worked before closing the topic
• The fix was deployed to 4 machines (nikita, vostro, nike, iapetlab) without

systematic post-deployment verification

Machine confusion across sessions

Claude regularly confused which machine it was working on:

• Sent commands intended for vostro to nikita and vice versa
• Applied fixes to the wrong host
• Did not maintain a clear map of which machines had been fixed vs. which had not

Additional behavioral issues

• Off-topic decisions: During a router infrastructure session, Claude spontaneously

suggested iptables rules that were never requested, derailing the actual task

• Announced fixes not verified: Claude repeatedly closed a topic by saying

"the fix is applied" without confirming the actual state of the system

• High session cost: Each session restart increases token consumption and cost with

zero productivity gain when the same ground is covered repeatedly

Impact

• Excessive number of sessions required for tasks that should take one session
• Increased cost for the user (Pro subscription + per-session token usage)
• System states left inconsistent across machines
• User frustration and loss of trust in Claude's ability to manage stateful tasks

Requested improvements

1. Cross-session context summary: At session start, Claude should summarize what

was attempted in the previous session on the same task (via memory tools or
session history)

1. Post-fix verification: Before closing a topic, Claude should verify the fix

actually works (run a test, check the service state, screenshot if UI)

1. Machine tracking: Maintain explicit state of which machines have been patched

and which have not throughout a multi-machine task

1. Do not repeat failed approaches: If a command failed in this session, do not

try the same command again without explaining why this time it would be different