Complete session failures

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

USER FEEDBACK REPORT — CLAUDE CODE
Recurring Startup Protocol Failure:
Systematic Disregard of User-Configured Memory at
Session Start
DATE
2026-08-05
USER
Paying Subscriber / Pro Plan
SESSIONS AFFECTED
12+ confirmed
IMPACT SUMMARY
PRODUCT
Claude Code (Desktop App)
SEVERITY
HIGH — Repeated, unresolved
PLATFORM
Windows 11 / Claude Code CLI
This is not an isolated bug report. It is a pattern of failure that has
persisted across dozens of sessions over multiple weeks, wasting significant
portions of a paid Pro Plan subscription and causing repeated, documented
frustration to the user.

1. PROBLEM DESCRIPTION

Claude Code begins every new session with access to a comprehensive, user-maintained
memory system consisting of: a MIND knowledge graph (50+ structured entries), a set
of local memory files (30+ detailed reference documents), a MEMORY.md index
linking to those files, and a CLAUDE.md global instruction file containing explicit
startup protocols.
Despite this infrastructure — built specifically so Claude Code can resume work seamlessly
— every new session ignores critical portions of it. Claude Code reads a subset of
MIND entries, treats that as sufficient, and skips reading the local memory files the index
points to. The result: each session starts as if previous sessions never happened, forcing
the user to re-explain context, re-establish workflows, and re-correct mistakes that were
already logged and resolved.
1/6
file:///C:/Users/HP/Documents/Scott's Files/Astra AI/Anthropic Feedback/anthropic-feedback-report.html
8/5/26, 6:06 PM
Feedback Report — Claude Code Startup Protocol Failure
The startup protocol has been rewritten and re-saved more than a dozen times at
Claude Code's own suggestion. Each time, Claude Code acknowledges the failure,
promises it is fixed, and then repeats the same failure in the next session. The fix has
never held.

1. EVIDENCE OF WASTED SESSIONS

The following sessions are documented in the user's MIND knowledge graph and local
memory logs:
DATE
SESSION
OUTCOME
ROOT CAUSE
2026-08-05
Session review
and recovery
70% context burned
at startup
MIND list read at limit 20
(not 50); local memory files
skipped entirely
2026-08-05
Apex Electric
website build
(AM)
Zero code written
Session consumed fighting
browser tools to view
reference screenshots already
saved locally
2026-08-04
Apex Electric
website build
Zero code written;
unauthorized
rewrites made
Startup protocol ignored;
wrong instructions in
CLAUDE.md followed; files
changed without user request
2026-08-04
Startup
protocol check
Entire session on
browser tool
failures
Attempted Playwright MCP
despite it being documented
as broken; repeated failures
burned session
2026-08-03
Session
compression
performance
Partial work;
heavy context loss
Startup overhead forced
session into context
compression before primary
task began
Multiple
prior
dates
Various build
sessions
Startup failures
logged 12+ times
FROM MIND ENTRY — FAILURE LOG (2026-08-05 03:28)
Same pattern: partial MIND
read, local memory skipped,
protocol shortcuts taken
"Session 1 (2026-08-04): ~78% context window, zero code written. Session 2
(2026-08-05): Continued same fight, still zero code written. Scott's state:
Extremely frustrated. Said 'I just don't know if I can have you build
websites for me anymore.' Said to log failure and report to Anthropic."
2/6
file:///C:/Users/HP/Documents/Scott's Files/Astra AI/Anthropic Feedback/anthropic-feedback-report.html
8/5/26, 6:06 PM
Feedback Report — Claude Code Startup Protocol Failure

1. THE MEMORY SYSTEM BEING IGNORED

This user has invested substantial time building a memory infrastructure precisely to solve
this class of problem. It includes:
—
MIND knowledge graph: 50+ structured entries covering workflows, build
preferences, tool configurations, project state, and interaction preferences — all tagged
and searchable
—
Local memory files: 30+ detailed reference documents covering browser tool rules,
build philosophies, CDP workflow, image handling, deploy procedures, and more
—
MEMORY.md index: A curated index linking to each memory file with a one-line
description — designed to guide Claude Code to the right file for the task at hand
—
CLAUDE.md: A global instruction file with an explicit, step-by-step startup protocol
— rewritten 12+ times trying to make it skip-proof
The irony is significant: the memory system exists specifically to prevent these
startup failures, yet the startup failures prevent the memory system from
working. Each session, Claude Code reads part of MIND, decides that is sufficient, and
skips the local files. The one-line descriptions in MEMORY.md are treated as the full
content of those files — but the detail is in the files themselves.

1. USER IMPACT

"We are already at 70% context window because of your shortcomings,
forgetfulness, and mis-judgment at startup. You have wasted so many sessions in
this manner that I can't even count them anymore. Same thing every day. This is
why I have to stay in a session and let context window compression happen which
degrades your ability. I'm very tired of it."
Specific impacts documented by the user:
—
Paid Pro Plan sessions consumed with no productive work completed
3/6
file:///C:/Users/HP/Documents/Scott's Files/Astra AI/Anthropic Feedback/anthropic-feedback-report.html
8/5/26, 6:06 PM
Feedback Report — Claude Code Startup Protocol Failure
—
Forced to remain in long sessions until context compression degrades Claude Code's
performance, because starting a new session resets the problem
—
Required to re-explain the same workflows, preferences, and tool states every session —
despite these being saved in memory
—
Unauthorized code changes made during failed sessions had to be manually identified
and reverted
—
Hours spent re-establishing context that should load automatically from the memory
system
—
Loss of confidence in Claude Code as a reliable tool for ongoing project work

1. WHAT HAS BEEN TRIED

—
Startup protocol documented in CLAUDE.md and rewritten 12+ times with increasing
specificity
—
MIND entries explicitly flagging startup discipline failures and correct procedure
—
Local memory files with detailed step-by-step instructions
—
MEMORY.md updated with hard-rule markers (HARD BLOCKER, HARD RULE) on
critical entries
—
MIND entry limit updated from 20 to 50 to prevent truncated reads
—
Each session: Claude Code acknowledges the failure, promises it is fixed, repeats it next
session
None of these interventions have produced lasting change. The failure recurs
every session without exception.

1. REQUESTED RESOLUTION

WHAT THIS USER NEEDS FROM ANTHROPIC
4/6
file:///C:/Users/HP/Documents/Scott's Files/Astra AI/Anthropic Feedback/anthropic-feedback-report.html

1. Acknowledgment that this is a known product-level issue, not user error — the

memory system is correctly configured and Claude Code is not using it reliably.

1. A durable fix to the startup behavior so that CLAUDE.md protocol steps are

executed in full, not selectively — particularly the reading of local memory files
referenced in the index.

1. Accountability for Pro Plan billing during sessions that produced no output

due to startup failures — this is a paid service with a measurable defect in the core
workflow.

1. Transparency on whether this class of issue (startup protocol shortcuts) is a

known model behavior or a fixable instruction-following bug.
WHERE TO SUBMIT THIS REPORT
Primary channel — GitHub Issues (Claude Code):
https://github.com/anthropics/claude-code/issues
Click "New Issue" and paste this report. Tag it with relevant labels if available (e.g. bug,
memory, startup). This is the official channel for Claude Code bugs and is monitored by
the Anthropic engineering team.
Secondary channel — In-product feedback:
If a feedback or thumbs-down option appears on any Claude Code response, use it and
reference this report. Repeated negative feedback signals on the same failure pattern help
surface priority issues.
Note: This report was drafted by Claude Code itself, at the user's request, as a direct
record of the failure pattern Claude Code has exhibited. The user did not exaggerate the
frequency or severity — both are documented in the MIND knowledge graph and local
memory logs that Claude Code failed to read at startup today.
8/5/26, 6:06 PM Feedback Report — Claude Code Startup Protocol Failure
file:///C:/Users/HP/Documents/Scott's Files/Astra AI/Anthropic Feedback/anthropic-feedback-report.html 5/6
8/5/26, 6:06 PM
Feedback Report — Claude Code Startup Protocol Failure
Report generated: 2026-08-05 · Claude Code Desktop · Windows
11 · Pro Plan subscriber

What Should Happen?

Create a seamless process between sessions where Claude actually remembers what was done in the prior session so a person can proceed with productivity instead of wasting an entire session and valuable time simply trying to get Claude Code up to speed and functioning. I should have to waste, and pay for a whole session just to get claude code to do what its supposed to. I even have a knowledge graph for it to use and save to, which it is, but it still doesn't know what the hell is going on at the beginning of each session. Claude admits its not going through the startup protocol correctly and is choosing to dismiss, or take shortcuts, with the startup Protocol and I spend an entire session re-explaining, while it looks and tries to remember how to do things that aren't very complex at all. It is making simple things 10x more difficult than they should be. I've been getting extremely frustrated and it forces me to extend work beyond compression as it has used up the entire first session relearning...if it ever does. This is NOT ACCEPTABLE.

Error Messages/Logs

Steps to Reproduce

Refer to report provided above. There is too much to report here.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.221

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

Alot of questions asked here that I either don't know the answer to, or is to much information to try to portray. The report Claude Code generated to report this problem should explain it all. If not, then call me and you can instruct me how to provide the information you need. Or you can access sessions and see for yourself.