VividnessMem beats Letta on memory benchmarks

My read: yes, this is probably doing less than Letta in the ways that matter for this benchmark, and that is the point. The public repo frames it as a smaller, opinionated memory policy rather than a general agent platform.

• –The repo describes a zero-dependency memory system with no embeddings and no vector DB, using lexical resonance, co-occurrence, and vividness scoring instead.
• –That kind of constrained retrieval can absolutely punch above its weight on Mem2ActBench, which rewards surfacing the right memory for tool selection more than broad semantic search.
• –Letta’s docs show a broader architecture: pinned core memory blocks plus vector-backed archival memory. More flexible, but also more moving parts.
• –The 22M figure should be treated as benchmark-specific until the exact config is public; the repo README I found documents a different 12B Gemma-based evaluation, so there may be multiple test setups in play.