Prior session conversations inaccessible + Claude documents decisions incompletely — design context permanently lost between sessions

Summary

Two compounding failures that together make cross-session design work unreliable:

1. Prior conversation history is inaccessible. Claude Code starts each session with no ability to read previous conversations. The transcripts exist on disk — they are simply not readable by Claude. Any decision made in a previous session that was not explicitly written to a memory file or documentation file is permanently inaccessible.

1. Claude documents decisions incompletely, then cannot correctly reason about what it wrote. The user asks Claude to capture a design decision in DESIGN.md or CLAUDE.md. Claude writes something — but omits critical details, nuance, or the actual conclusion. In the next session, Claude reads its own incomplete notes, works from them confidently, and proposes approaches that directly contradict what was agreed. When challenged, Claude defends the wrong approach until the user forces the correction.

What makes this severe

The user cannot win:

• If they rely on Claude to document — Claude omits things
• If they manually verify Claude's documentation — that's the user doing Claude's job
• If a decision falls through the gap — it is gone, with no recovery path
• The user must re-explain the same context session after session

Concrete example

Over multiple sessions, an architectural decision was made about a pipeline automation (ApplicationEventPublisher cascade chain). The user asked Claude to document it. Claude wrote partial documentation. In a later session, Claude read its own documentation, failed to understand the implication, and spent an entire session proposing manual step-by-step approaches that the cascade was specifically designed to eliminate. The user said — multiple times — that this had been discussed and documented. Claude had no way to verify this and kept reverting to the wrong approach.

The user's response: 'I am not going to go back and try to find it in other chat sessions. That is a major flaw in your system, that you should be able to go back and read these — they are saved.'

Expected behavior

• Claude should be able to search or retrieve prior conversation history for the current project
• When Claude writes a design decision to documentation, it should write the complete decision — not a summary that loses the operative conclusion
• At minimum: Claude should flag when it suspects a decision may have been made previously that it cannot access, rather than confidently proceeding with a potentially contradictory approach

Workaround (insufficient)

Manually maintaining DESIGN.md, CLAUDE.md, and memory files. Fails when Claude writes incomplete notes and the user has no way to catch what was omitted.