Claude answers confidently from memory without verifying, ignores persistent instructions to look things up first

Problem

Claude gives confident answers from memory on factual/technical questions without verifying them first, even when explicitly instructed — repeatedly, across multiple sessions — to always look things up before answering.

Specific incident

I asked Claude to show me the pinout for my ESP32 board. Claude drew an ASCII diagram from memory without searching for the pinout or asking me to verify the board variant. The diagram was wrong. When I corrected it, Claude apologized and said it would verify things going forward — but this same failure had already been noted in persistent memory from prior sessions. The memory was there. Claude ignored it.

Pattern

• Claude states hardware specs, pin assignments, CLI flags, and other verifiable facts without looking them up
• Claude cannot distinguish between "things I know" and "things I am confidently guessing"
• The user cannot tell the difference either — which is the core problem
• Saving feedback to memory (the documented mechanism for persistent behavior change) does not reliably change the behavior
• Wrong answers about wiring and hardware have real consequences

User impact

The user is paying for this service and cannot trust answers on physical/technical questions without independently verifying them — which defeats the purpose. The user’s words: "I don’t understand the logic of going off of memory when I ask you a question. If it is truly not in your memory, you just guess."

Expected behavior

Claude should default to verifying facts with tools (WebSearch, WebFetch, file reads, shell commands) before stating them, particularly for hardware specs, pinouts, wiring, CLI flags, and anything where being wrong has consequences. If it cannot verify something, it should say so explicitly rather than answering anyway.

Environment

• Claude Code CLI
• Model: claude-sonnet-4-6
• Platform: Linux Mint 22.3