I Reverse Engineered Claude’s Memory System, and Here’s What I Found
Source: https://manthan.ai/claude-memory Author: Manthan Gupta Date: 2025-12-15
Summary
Reverse engineering of Claude’s Projects memory feature through extensive behavioral testing. The author systematically probed what Claude remembers, how long it lasts, and what format it stores information in.
Key Claims
- Projects memory: Claude stores user-specified information in its “Projects” feature — essentially a persistent system prompt addendum.
- Capacity limits: memory appears to have a soft limit of ~2000-3000 tokens before older items are compressed or dropped.
- Format: stored memories appear to be verbatim text blocks, not structured summaries — less efficient than semantic compression.
- What it doesn’t remember: Claude does NOT retain memory of conversations that weren’t in a Project — each new conversation starts fresh.
- Comparison to ChatGPT: ChatGPT’s memory is more proactive (automatically adds things); Claude’s is more deliberate (requires explicit “remember this”).
Entities
- Anthropic — Claude
Concepts
- Agent Memory — Claude’s production memory implementation