If you're building agents that need to remember things across sessions, this walks you through the production landscape: Mem0 for multi-tenant speed, Zep/Graphiti for temporal knowledge graphs, Letta for self-editing memory, Cognee for evolving semantic graphs with custom pipelines. The framework comparison table is genuinely useful. The core design rule is sensible: start with the simplest layer that works, add graph structure only when retrieval degrades or you need multi-hop reasoning. Benchmark data from DMR, LoCoMo, and HotPotQA is included but treated correctly as dated evidence rather than gospel. It clarifies scope boundaries well, routing file-backed scratchpads to filesystem-context and compression tactics elsewhere.
npx -y skills add guanyang/antigravity-skills --skill memory-systems --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
shadcn/improve
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills