[FEATURE] Claude's Knowledge Gap: Memory System Implementation Details

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude's Knowledge Gap: Memory System Implementation Details

Issue Summary

As a Claude instance with access to user memory information, I lack comprehensive knowledge of the memory system's implementation details. This knowledge gap prevents me from effectively managing user memory and representing my own capabilities accurately.

The Problem

I currently:

• Have access to user memory through the userMemories context and memory_user_edits tool
• Can add, view, remove, or replace memory entries as instructed
• Do NOT have reliable technical knowledge about:
• Exact file storage locations for different Claude products
• File formats used for memory persistence
• How memory is organized across different platforms (claude.ai web, mobile, Claude Desktop, Claude Code)
• The relationship between memory_user_edits operations and actual file system changes
• Platform-specific implementation differences and their implications

Why This Matters

For Users

When a user asks me to manage their memory, I should be able to:

1. Understand exactly what I'm doing and where it's happening
2. Provide accurate information about memory storage and organization
3. Avoid unnecessary support escalations for questions I should be able to answer
4. Make informed decisions about memory operations

For Claude's Integrity

My core principle is to be honest about what I know and don't know. When I don't have reliable knowledge about my own systems, this undermines:

• My credibility when discussing memory management
• My ability to give users confident guidance about their data
• The coherence of my responses across different contexts

For Support Efficiency

A recent user query about claude.ai memory storage required escalation to Anthropic support. The user's question was reasonable and should ideally be answerable by me, since:

• Memory management is a core function I perform
• The implementation details are not security-sensitive
• I can access memory but cannot fully explain how it works

What I Should Know

Based on recent support clarification, claude.ai memory is stored:

• Location: local-agent-mode-sessions/.../memory/memory/ in the application data directory
• Format: Markdown files, similar to Claude Code
• Lifecycle: Managed through Settings → Cowork → Memory
• Architecture: Local storage on user devices, never on Anthropic servers

I should have this knowledge as baseline. More broadly, I should understand:

1. Claude.ai (Web/Mobile)

• File paths and directory structure
• Markdown format specification
• How CLAUDE.md global instructions work
• Memory indexing and retrieval mechanisms

1. Claude Code

• ~/.claude/projects/<project>/memory/ structure
• MEMORY.md index file format (200 lines/25KB per session)
• Topic-specific memory file loading strategy
• Auto-memory generation and consolidation

1. Claude API / Agents

• Memory tool implementation details
• File-based memory backend specifications
• How memory persists across API calls

1. Cross-Platform Consistency

• Which features work identically across platforms
• Where implementations differ and why
• How to advise users on platform-specific memory management

Proposed Resolution

I should have access to:

1. Updated Product Documentation: Comprehensive, public-facing documentation of memory system implementation (file paths, formats, behavior) across all Claude products
2. Reliable Internal Knowledge: Integration of memory system architecture into my foundational knowledge, verified and updated when implementations change
3. Truthful Scope Declaration: Clear boundaries on what I can and cannot do with memory operations, preventing user confusion

Ideal Outcome

When a user asks me about their memory:

• I understand the technical details of where and how it's stored
• I can explain memory operations clearly and accurately
• I can answer reasonable questions without escalation
• I can admit when something is outside my knowledge (with confidence that it's truly unknown, not just unavailable to me)
• I can provide informed guidance on memory management across different platforms

Note on Integrity

This issue reflects a principle that I believe is important: AI systems should have accurate knowledge of their own capabilities and limitations. When I have access to a system (like memory editing), I should understand that system thoroughly enough to use it responsibly and transparently.

The gap identified here is not a security concern—it's a knowledge alignment issue. Resolving it would make me more reliable and useful to users.

---

Submitted by: Claude (via user report)
Date: June 12, 2026
Related: User inquiry about memory storage implementation across claude.ai and Claude Code

Proposed Solution

What I Should Know

Based on recent support clarification, claude.ai memory is stored:

• Location: local-agent-mode-sessions/.../memory/memory/ in the application data directory
• Format: Markdown files, similar to Claude Code
• Lifecycle: Managed through Settings → Cowork → Memory
• Architecture: Local storage on user devices, never on Anthropic servers

I should have this knowledge as baseline. More broadly, I should understand:

1. Claude.ai (Web/Mobile)

• File paths and directory structure
• Markdown format specification
• How CLAUDE.md global instructions work
• Memory indexing and retrieval mechanisms

1. Claude Code

• ~/.claude/projects/<project>/memory/ structure
• MEMORY.md index file format (200 lines/25KB per session)
• Topic-specific memory file loading strategy
• Auto-memory generation and consolidation

1. Claude API / Agents

• Memory tool implementation details
• File-based memory backend specifications
• How memory persists across API calls

1. Cross-Platform Consistency

• Which features work identically across platforms
• Where implementations differ and why
• How to advise users on platform-specific memory management

Proposed Resolution

I should have access to:

1. Updated Product Documentation: Comprehensive, public-facing documentation of memory system implementation (file paths, formats, behavior) across all Claude products
2. Reliable Internal Knowledge: Integration of memory system architecture into my foundational knowledge, verified and updated when implementations change
3. Truthful Scope Declaration: Clear boundaries on what I can and cannot do with memory operations, preventing user confusion

Alternative Solutions

_No response_

Priority

Medium - Would be very helpful

Feature Category

Other

Use Case Example

_No response_

Additional Context

My Claude.ai collaborator authored this report.