anthropics / anthropics/claude-code
[BUG] Claude Code: Memory/Standing Instructions Inconsistently Applied Over Multiple Sessions
- Dominant language
- Python
- Stars
- 145k
- Forks
- 23.1k
- PR merge metrics
- PR metrics pending
Description
### Preflight Checklist
- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) 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?
## Claude Code: Memory/Standing Instructions Inconsistently Applied Over Multiple Sessions
### Description
Over the past week, Claude has been inconsistently reading and following standing instructions stored in the memory system (`~/.claude/projects/c--Repos/memory/`), despite having explicit access to them. This has resulted in repeated errors that directly violate documented feedback and conventions.
### Expected Behavior
When Claude encounters a task that matches a remembered instruction or past feedback (e.g., "save documents to Work Notes folder X, not repo folder Y"), Claude should:
1. Read the relevant memory file before acting
2. Apply the instruction consistently
3. Not require repeated corrections for the same instruction
### Actual Behavior
Claude is skipping memory checks before taking action, leading to:
- Files saved to wrong locations despite explicit memory feedback
- Product-specific work done without checking stored conventions
- Repeated mistakes on tasks that previous memory indicates were already corrected
- Pattern has been consistent across multiple separate tasks over ~7 days
### Steps to Reproduce
1. Set up a memory system with feedback on a common task type (e.g., document locations)
2. Ask Claude to perform that task in a new conversation
3. Observe: Claude does not read the memory before acting
4. Correct Claude; the correction is acknowledged but same error repeats in subsequent tasks
### Example
- Memory file: `feedback_docs_go_to_work_notes.md` explicitly lists where each product's docs should be saved
- Task: Create a KB article for NAA
- Observed: Claude created file in new `Products\NAA` directory
- Expected: Claude would read memory first, find `2. Access Analyzer (Classic)\2b. Docs`, save there
- Result: File created in wrong location; user had to intervene
### Environment
- Claude Code (VSCode Extension)
- Session span: Sept 11-18, 2026
- Multiple separate conversations
### Additional Context
Neither the user nor Claude can identify what changed or why this pattern emerged. The memory system is present and accessible, but not being reliably consulted before action. This suggests a systematic issue with how standing instructions are being processed, not a user error or configuration issue.
### Impact
Repeated wasted time on corrections that were already explicitly documented. Standing instructions become unreliable, defeating the purpose of the memory system.
### What Should Happen?
It should be reviewing the memory files, otherwise what is the point of them?
### Error Messages/Logs
```shell
```
### Steps to Reproduce
1. Set up a memory system with feedback on a common task type (e.g., document locations)
2. Ask Claude to perform that task in a new conversation
3. Observe: Claude does not read the memory before acting
4. Correct Claude; the correction is acknowledged but same error repeats in subsequent tasks
### Claude Model
Sonnet (default)
### Is this a regression?
Yes, this worked in a previous version
### Last Working Version
_No response_
### Claude Code Version
2.0.24 (Claude Code)
### Platform
Anthropic API
### Operating System
Windows
### Terminal/Shell
VS Code integrated terminal
### Additional Information
_No response_
Contributor guide
No contributing guide indexed for this repository
Research direction
Start by reproducing the behavior in the VS Code extension with the memory file at ~/.claude/projects/c--Repos/memory/feedback_docs_go_to_work_notes.md and a new conversation. Compare whether Claude reads the memory before handling the document-location task across sessions; done means the remembered destination is consistently applied without repeated correction.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- vscode
- Domain
- tooling
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100