Skip to content

Memory

The fork adds a cross-session memory system, enabled by default. The agent can consult what it learned before and save new facts for later — across sessions and across projects.

Scope Location Availability
Session Current session Immediate context
Project ~/.local/share/opencode/memory/projects/<project>/ Every session in that project
Global ~/.local/share/opencode/memory/global/ Every session, every project

Global memory is also available as an automatically loaded skill, and both scopes can be queried on demand by the model through dedicated tools — it searches when relevant instead of flooding every prompt.

A native memory tab sits in the sidebar next to Context, with:

  • Session/Project and Global subtabs
  • Real-time search across entries
  • Semantic badges distinguishing decisions from facts
  • Backfill — retroactively extract memories from past sessions

Memory synthesis uses an intelligent multi-provider fallback chain (AGY / OpenRouter / Zen / heuristic). If you don’t pick a model explicitly, memory automatically uses the default system/chat model — including combos and connected models.

Enable, disable and choose the memory model in Settings → Memory, with the same instant-toggle pattern as the rest of the app. Memory entries can also be read, added and promoted through the HTTP memory API — per project and globally.