Most agents start every session amnesiac: you either re-paste history into the context window (expensive, lossy) or bolt on a vector store and hope retrieval surfaces the right thing. Memori reframes memory as structured state rather than a pile of retrievable chunks — it parses conversations and tool calls into typed facts, events, preferences, relationships, and skills, so recall is targeted instead of a fuzzy similarity search.
The number that makes the case: on the LoCoMo benchmark it reports 81.95% accuracy at an average of 1,294 tokens per query. That second figure is the real point — instead of stuffing the whole transcript back in, it injects only the slice that matters, which is what keeps long-running agents both accurate and cheap.
What Sets It Apart
- Memory is typed, not just embedded — attributes, events, facts, relationships, preferences, and skills are tracked distinctly, so you can reason over what kind of thing was remembered, not just nearest-neighbor text.
- Genuinely LLM-agnostic: works across OpenAI, Anthropic, Bedrock, Gemini, DeepSeek, and Grok, so memory isn't coupled to one provider.
- Two ways in — direct SDK (Python and TypeScript) for app builders, or MCP for drop-in persistence in Claude Code, Cursor, and similar clients without writing integration code.
Who It's For
A strong fit if you're shipping production agents that must remember users or accumulate knowledge across sessions, and you care about token cost at scale. Look elsewhere if you need a single-shot chatbot with no continuity, or if your memory needs are simple enough that a basic vector store and a system prompt already cover them — Memori's structured-state model is overhead you won't recoup at that size.
