OpenCode Needs Memory, Not Another Background Pipeline

OpenCode's new plugin memory aims to address the issue of agents forgetting valuable lessons learned across sessions. Existing solutions like AGENTS.md and session history have limitations - AGENTS.md is great for defining rules but not for retaining reusable knowledge, while session history doesn't automatically surface relevant information across sessions.

The plugin introduces four memory tools: memory_add to save reusable project facts or conventions, memory_update to correct or expand existing memories, memory_delete to archive obsolete entries, and memory_read to retrieve full entries with optional filtering. These tools give the active agent control over what knowledge to retain.

The memory system stores knowledge as project-local Markdown files under .opencode/memory/, with MEMORY.md as the source of truth that can be edited directly. A generated memory_summary.md provides a concise index of active memory entries to include in system prompts, enabling ambient awareness of relevant memory without loading the entire handbook each time.

The plugin maintains a plugin.log for tracking memory modifications, offering transparency into how the memory layer evolves. Markdown is chosen as the storage format because it gives users direct ownership of the memory files, makes operations visible, preserves stable IDs, allows soft deletions by archiving entries instead of erasing them, and maintains a clear audit trail.

By focusing on selective recall and providing tools for deliberate memory promotion at the moment knowledge is fresh, the plugin aims to create a memory layer owned by the active agent, preserving valuable lessons without the noise of recording every task.