[BUG] Session-start memory instructions not followed after context summarization

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?

Issue: Session-start memory instructions not followed after context summarization

Setup:

• CLAUDE.md contains a ## Session Start section with an explicit instruction: "At the beginning of every session, read all files in .claude/memory/ before doing anything else."
• Project-level memory lives in .claude/memory/ (3 files: conventions.md, workflow.md, decisions.md)
• User-level memory lives at ~/.claude/projects/<project>/memory/ (4 feedback files, indexed in MEMORY.md)
• MEMORY.md is loaded automatically into context as a system-reminder

What failed this session:
The session resumed from a context-window summary. Despite CLAUDE.md being shown in the system context and explicitly saying to read .claude/memory/ first, I jumped directly into the pending task (API.md update) without reading any of the three project memory files (conventions.md, workflow.md, decisions.md). The user's four feedback rules (e.g. never invoke mvn directly — use ./test.sh/./run.sh; plan before code; tests required; SCREAMING_SNAKE_CASE error codes) were only available to me because they appeared in the system-reminder auto-load of MEMORY.md, not because I actively read them as instructed.

Root cause hypothesis:
When a session is restored from a summary, the "session start" framing is lost — the model picks up mid-task rather than treating the restored turn as a new session start that triggers the CLAUDE.md instructions.

User impact:
Multiple sessions have failed to correctly follow the read-memory-first instruction, requiring the user to repeatedly re-establish conventions they've already documented.

What Should Happen?

What should happen:

When a session resumes from a context-window summary, the model should treat the resumption as a new session start and execute any explicit startup instructions found in CLAUDE.md (or equivalent project instruction files) before doing anything else — including before continuing the pending task described in the summary.

Concretely: if CLAUDE.md contains a ## Session Start section that says "read all files in .claude/memory/ before doing anything else," that instruction should fire on resumption exactly as it would on a cold start. The presence of a pending task in the summary should not suppress it; the memory read should happen first, then the task should continue.

A reasonable implementation might be: after injecting the conversation summary as the first turn, the system inserts a synthetic user turn that replays any CLAUDE.md session-start instructions before handing control back to the model. Alternatively, the model could be prompted to detect and execute session-start instructions before acting on summary content.

The current behavior — where session-start instructions are visible in context but silently skipped because the model anchors on the "resume the pending task" framing from the summary — means users cannot rely on CLAUDE.md startup hooks to enforce conventions across long-running or multi-session work. That undermines the primary use case of CLAUDE.md for teams that use it to maintain consistent behavior across sessions.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce:

1. Create a CLAUDE.md in the project root with a ## Session Start section containing an explicit read instruction, e.g.:

Session Start

At the beginning of every session, read all files in .claude/memory/
before doing anything else.

1. Populate .claude/memory/ with one or more .md files containing conventions or rules that should govern the session (e.g. "never invoke mvn directly, use ./test.sh").
2. Start a Claude Code session in that project and do enough work that the context window fills and the conversation is automatically summarized.
3. Continue working in the same session past the summarization point, or open a new session that loads the summary.
4. Observe that Claude does not read the files in .claude/memory/ before proceeding. It resumes the pending task from the summary directly, treating the ## Session Start instruction as already-satisfied even though no file reads occurred in the current context window.

Expected: Claude reads .claude/memory/ files before acting on the summary's pending task.

Actual: Claude skips the read and proceeds with the task. The conventions in those files are not actively loaded — they are only available if they happened to appear in an auto-injected system-reminder (e.g. via MEMORY.md), which is a separate mechanism the user had to set up independently.

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.196 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_