Pick the edition that fits your stack. Both ship the same audit-grade cognitive memory engine. Studio adds Vexilon, our bundled hybrid retrieval layer, for teams who want corpus search in the box.
What your agent remembers. STARE 5D scoring, deterministic compile, every entry traced to source.
Everything in Engine, plus Vexilon, our production hybrid retrieval layer with knowledge graph and reranking.
Every capability either ships with both editions, or is exclusive to Studio.
| Capability | Engine | Studio |
|---|---|---|
| Agent coordination | ||
| Switchback agent coordination MCP server | ✓ | ✓ |
| Cognitive memory | ||
| STARE 5D scoring | ✓ | ✓ |
| Auto-STARE with bundled local embeddings + distilled Significance scorer (CPU; extraction-grade scoring uses your LLM) | ✓ | ✓ |
| Memory Consolidation (5-step pipeline) | ✓ | ✓ |
| Episode summaries (stable) | ✓ | ✓ |
| Narrative clustering (beta, opt-in) | ✓ | ✓ |
| R-graph relationships | ✓ | ✓ |
| Salience scoring & decay | ✓ | ✓ |
| Memory Seeding (bulk import for cold-start onboarding) | ✓ | ✓ |
| Backends & deployment | ||
| File backend (.md-native, agent-isolated) | ✓ | ✓ |
| Postgres backend (multi-tenant, SQL audit) | ✓ | ✓ |
| Idempotency primitives (Postgres) | ✓ | ✓ |
| Helm chart & Docker Compose | ✓ | ✓ |
| Licensed self-host (you run it) | ✓ | ✓ |
| Audit & provenance | ||
| Provenance chain on every memory | ✓ | ✓ |
| Audit hash-chain (Postgres) | ✓ | ✓ |
| OTEL-native tracing | ✓ | ✓ |
| Developer experience | ||
| MCP tools (mm_remember, mm_search, ...) | ✓ | ✓ |
| Python SDK | ✓ | ✓ |
| TypeScript SDK | roadmap | roadmap |
| Framework adapters | ✓ | ✓ |
| BYO-RAG adapter (LlamaIndex, LangGraph, Vespa) | ✓ | ✓ |
| Vexilon retrieval layer (Studio only) | ||
| Hybrid retrieval (dense + BM25, RRF fusion) | — | ✓ |
| Cross-encoder reranker (BGE-reranker-v2-m3) | — | ✓ |
| Contextual chunking (LLM-generated preambles) | — | ✓ |
| Knowledge graph (entities + relationships) | — | ✓ |
| Boost profiles (recent / significant / procedural) | — | ✓ |
| Filesystem watcher + git post-commit hooks | — | ✓ |
| Incremental reindex API | — | ✓ |
| Bundled vector store (Qdrant) | — | ✓ |
| Bundled embedding service (BGE-M3) | — | ✓ |
| Unified MCP tools across memory and corpus | — | ✓ |
Retrieval-augmented generation answers “what does the document say?” Cognitive memory answers “what did we decide, and why?” Most production agent fleets need both. We bundle them only when you want them bundled.
MM Engine is for teams who already have a RAG stack they like (LlamaIndex, LangGraph, Vespa, pgvector, custom). The Engine ships with an adapter so retrieved context flows cleanly into the consolidation pipeline. You keep your retrieval layer. We add the memory layer.
MM Studio is for teams who’d rather not wire their own. Studio includes Vexilon, our production retrieval layer (hybrid dense + sparse search, cross-encoder reranking, contextual chunking, knowledge graph). One install. One MCP surface. Memory and corpus together.
Both editions are licensed self-host. You run the binary in your infrastructure. We never see your data. Same product cadence, same audit-grade provenance guarantees.
Request a license and tell us about your stack. We’ll recommend the right edition and walk you through the install path for your deployment shape, Docker Compose, Kubernetes via Helm, or bare metal.
Request a license