Auto Memory

OpenOrca can automatically learn and remember project-specific patterns, conventions, and insights across sessions. At the end of each session, the LLM summarizes what it learned and saves it for future reference.

How It Works

1. During a session: You interact with OpenOrca normally, using tools to read/write code, run commands, etc.
2. At session end: If auto memory is enabled and the session involved meaningful tool usage, OpenOrca asks the LLM: "What project-specific patterns, conventions, or learnings should be remembered?"
3. Save: The LLM's summary is saved as a .md file in the memory directory
4. Next session: All memory files are loaded into the system prompt, giving the LLM context about your project from previous sessions

Memory Directories

Location	Scope	Priority
.orca/memory/	Project-level	Loaded first
~/.openorca/memory/	Global	Loaded second

Project memory is loaded before global memory. When saving, learnings go to the project memory directory if a project root is found, otherwise to the global directory.

Configuration

In ~/.openorca/config.json:

{
  "memory": {
    "autoMemoryEnabled": true,
    "maxMemoryFiles": 20
  }
}

Setting	Type	Default	Description
autoMemoryEnabled	bool	true	Save session learnings automatically at session end
maxMemoryFiles	int	20	Maximum memory files per directory. Oldest files are pruned when exceeded.

Commands

/memory list

Show all auto-generated memory files with first-line previews.

> /memory list

/memory auto on|off

Toggle auto memory on or off for the current configuration.

> /memory auto on    # Enable auto memory
> /memory auto off   # Disable auto memory

/memory clear-auto

Delete all auto-generated memory files from both project and global directories.

> /memory clear-auto

/memory show / /memory edit

These existing commands still work for viewing and editing the ORCA.md project instructions file (separate from auto memory).

File Format

Memory files are named {date}-{hash}.md (e.g., 20260221-a1b2c3.md). The hash is derived from the content to avoid duplicates.

Example memory file content:

- Project uses file-scoped namespaces and sealed classes by default
- Tests are in xUnit with IDisposable temp directories for isolation
- Config is loaded from ~/.openorca/config.json via ConfigManager
- Tools implement IOrcaTool and are auto-discovered by ToolRegistry

System Prompt Integration

Memory content is injected into the system prompt under a PROJECT MEMORY (auto-learned): section, after the project instructions (ORCA.md). This gives the LLM persistent context about your project without you having to repeat information.

Tips

• Auto memory works best with consistent project usage over multiple sessions
• Project-level memory (.orca/memory/) can be added to .gitignore or committed for team sharing
• Use /memory clear-auto if memory becomes stale or incorrect
• Global memory captures patterns useful across all projects (e.g., your preferred coding style)
• The maxMemoryFiles limit prevents unbounded growth — oldest files are automatically pruned