The Memory System

Large language models have a fundamental limitation: every new conversation starts from scratch. They don't remember what you told them yesterday, don't know your preferences, and have no context about your project. This means you might need to repeatedly explain: • What tech stack your project uses • What coding style you prefer • What bug was fixed in the last conversation • Your role and work priorities The memory system exists to solve this problem.

OpenHarness uses a file-based memory system with these core components: 📄 MEMORY.md The memory index file that lists all memory entries with titles and links. It's loaded into context at the start of every conversation. 📁 Memory Directory Each memory entry is stored as an independent Markdown file, organized by topic. The design is elegant: MEMORY.md serves as an index that's always in context, so the Agent quickly knows what memories are available. When needed, it reads the specific files.

OpenHarness defines four memory types: 👤 User Memory Information about the user: role, preferences, expertise level, etc. Example: "User is a senior Go developer but new to React" 💬 Feedback Memory User corrections and confirmations of Agent behavior. Example: "Don't summarize at the end of every response" 📋 Project Memory Project-related information: goals, deadlines, special agreements. Example: "Code freeze starts March 5th" 🔗 Reference Memory Locations of external resources. Example: "Bug tracking is in the Linear INGEST project"

1. Create: The Agent writes a memory file when it discovers information worth remembering 2. Index: The memory entry is added to the MEMORY.md index 3. Load: MEMORY.md is loaded into context at the start of every new conversation 4. Use: The Agent adjusts its behavior based on memories 5. Update: If information becomes outdated, the Agent updates or deletes the memory Importantly, the Agent verifies that memories are still accurate before acting on them. Memories can become stale, so the Agent cross-references them with the current code state.