Claude loses in-session context at low utilization, forgets recent decisions and instructions

Problem

Claude Code loses track of information discussed earlier in the same session, even when context utilization is well under 40% of the 1M token window. This is not a context overflow issue — it happens consistently at low utilization.

Types of context loss observed:

• Forgetting which machine/server it's operating on
• Forgetting URLs provided earlier in the session
• Forgetting design decisions agreed upon minutes ago
• Forgetting the user's stated preferences/methodology
• Forgetting features that were already built

Evidence from session logs (last 7 days)

12+ incidents across 5 projects:

1. TEREDACTA (c8737c3c): Claude was supposed to be operating on a remote Linux VPS via a teleport session. Instead, it installed software on the local Mac under /Users/brianhill/. When confronted: "Whoa...hang on. Where did you do all this shit? Looks like you did it on my mac by the paths." The project memory file literally says "This project context is VPS admin, not app development."

1. TEREDACTA (d2fb8fd2): User gave a stress test URL at the start of the session. Claude later tested against the wrong server. User: "No, you tested against the wrong server. Can you even open the URL I gave you?"

1. Skippr (db9279a0): User: "again, agents should not need to ask to commit." Claude asked for commit permission despite the user having already told it agents should commit without asking.

1. Skippr (c8f141b6): User: "Again, if possible, wargame and test. I prefer this method... so long as we can do so quickly." Claude asked a multiple-choice design question instead of testing — user had already stated their preferred methodology.

1. Skippr (255d91ca): User: "I asked you, before you ever wrote a word of a spec, to move everything to the Skippr directory." Claude had put files in the wrong directory, forgetting a prior instruction.

1. Music-Discovery (c4db3873): User: "I could have sworn we discussed using playcounts to modify scoring, we even discussed how we wanted it to be logarithmic." Claude was treating play count as binary (played/unplayed).

1. Music-Discovery (c4db3873): User: "I have a musickit auth on my dev account with a key. We already use it for something in this project." Claude asked if the user had set up Music auth.

1. Music-Discovery (f7737d09): User: "I thought we already did a similar artists search for last.fm?" Claude proposed it as a new idea.

1. Music-Discovery (f7737d09): "Do these approaches take into account the signals we discussed earlier (feedback collection)?" Claude's proposals omitted previously discussed feedback signals.

1. MS-Clone (94723452): Claude was asked to install Chrome, installed chromium instead, edited a JSON config, told user to restart. When problem recurred, started doing the exact same wrong thing again. User: "You are fucking doing it again... Did you install the right version of chrome?"

Expected behavior

At <40% context utilization, Claude should reliably remember:

• Which machine/environment it's operating on
• URLs and endpoints provided in the current session
• Design decisions made in the current session
• User's stated methodology preferences
• What has already been built/discussed

Environment

• Claude Code CLI, macOS
• Context utilization consistently <40% of 1M tokens
• Happens across all projects, not project-specific