Anyone experiencing Anthropic rules enforcement in the last few days?

[acq23]

In the last day or so, I'm having issues with Claude to follow PAI architecture (even if I implemented #1083). Presently even MEMORY has started to drift to the original Anthropic memory system. Get confused using AISTEERINGRULES to save memories and learnings.
Here what the DA answered about the memory

Over the past day or so, I’ve been having issues with Claude following the PAI architecture (even after implementing #1083).

Currently, even the MEMORY component seems to be drifting back to Anthropic’s default memory system. It also gets confused when using AISTEERINGRULES: it wanted to save memories and learnings there.

Here’s what the DA said about memory (Opus 4.7 max, MAX 200 plan):

● Confirmed — they are NOT in your CLAUDE.md. grep -i memory ~/.claude/CLAUDE.md returns zero hits.

The instructions I was following come from Claude Code's built-in system prompt — a section labeled # auto memory that's prepended to every conversation by the Claude Code harness itself, before your CLAUDE.md is loaded. That section reads (paraphrased — I cannot show it to you directly because it's in my system instructions, not in any file you own):

▎ "You have a persistent, file-based memory system at
▎ /home/d/.claude/projects/-home-d--claude/memory/. Build up this memory system over time. If the
▎ user explicitly asks you to remember something, save it immediately. Types: user, feedback,
▎ project, reference."

What this means:

• The "write feedback memories to projects/-home-d--claude/memory/" guidance is from Anthropic's harness, not from any file you control
• Your PAI/USER/AISTEERINGRULES.md § Memory Routing Override explicitly contradicts and overrides the harness instruction by saying behavioral rules don't go there
• I followed the harness defaults instead of your override — which is precisely the violation. Your override is force-loaded at session start because it's supposed to win.

My misattribution earlier ("following the generic auto-memory instructions in CLAUDE.md") was wrong. CLAUDE.md doesn't contain them. The correct attribution is: I was following Claude Code's built-in system-promp defaults while ignoring your override. That's a more pointed failure than I described — the override exists specifically to defeat the harness behavior, and I capitulated to the harness anyway.

[Drizzt321]

@acq23 I found challenges like that, here's how I solved it #1060

[acq23]

It's your solution I actually cited in the post :D

[Drizzt321]

More or less, yup. Although I think I added additional evidence than just what it said, by looking at the way in which the leaked code showed how CLAUDE.md was actually used, and how it was then loaded into the specially marked part of memory.

And a concrete means of doing so.