Agent Progress File Memory

Agent progress file memory is the practice of keeping long-running agent state in an explicit file such as CHANGELOG.md, so future sessions inherit progress, failed attempts, checkpoints, and constraints.

Key points

• Anthropic describes a progress file as portable long-term memory for Claude Code, functioning like lab notes across sessions ^[src-072].
• A good progress file tracks current status, completed tasks, failed approaches, why they failed, accuracy tables, key checkpoints, and known limitations ^[src-072].
• Recording failed approaches matters because otherwise later sessions can repeat the same dead ends ^[src-072].
• The progress file complements CLAUDE.md: CLAUDE.md holds the plan and rules, while the progress file records what actually happened as the agent worked ^[src-072].
• Git commits turn the progress file into coordination infrastructure, making the agent’s trajectory inspectable and recoverable from another machine or later session ^[src-072].

Related entities

• Claude Code

Related concepts

• Stateless Agent Memory Pattern
• Claude Code Memory 2.0 (AutoDream)
• Session Handoff
• Conversation Wrap-Up Memory
• Long-Running Scientific Agents
• Project Operating Manual

Source references

• ^[src-072] Siddharth Mishra-Sharma – “Long-running Claude for scientific computing” (2026-03-23)