OpenZero is a substantial engineering effort on top of the OpenCode baseline. Here is exactly what changed.
| Capability | OpenCode | OpenZero |
|---|---|---|
| Memory system | None | Six typed schemas with custom extraction pipeline and mem0 integration |
| Memory storage | None | Qdrant vector database with 4096-dim embeddings |
| Deduplication | None | mem0-delegated deduplication via embedding similarity |
| Context management | Single-pass LLM compaction on overflow | Unified three-tier hierarchical compression (50/30/20 budget split) |
| Embedding support | None | Qwen3-Embedding-8B at 4096 dimensions via OpenRouter |
| Memory retrieval | None | Cosine-similarity semantic search with type-aware formatting |
| Configuration | JSON files (opencode.json / config.json) | JSON files plus SQLite for session and compression history |
Architectural decisions made in week 1 are still retrievable in week 3. No more re-explaining why you chose that database or that API pattern.
When you fix a bug, the symptom, root cause, and solution are extracted and stored. The same bug will never be re-investigated from scratch.
Hierarchical compression means you can have effectively unlimited conversation length. Recent context stays full fidelity while older history is summarized progressively.
Metadata-filtered search (by type, time range, trigger). Memory importance scoring. Analytics dashboard for retrieval effectiveness.
Memory relationships and dependency graphs. Advanced semantic summarization. Plugin marketplace. Multi-user memory sharing.
Federated memory across instances. ML-based ranking from retrieval feedback. Real-time collaboration. Enterprise features (SSO, audit logging, RBAC).
OpenZero installs in under a minute and runs entirely on your machine.