YourMemory shifts from flat RAG to graphs

Flat decay is too crude for agent memory: if a weak node is the only bridge to a strong one, pruning it can collapse the whole context graph. The graph-first move is the right direction, but the real challenge is balancing local relevance, structural importance, and privacy boundaries without turning recall into a noisy hub problem.

• –2-hop BFS is a sensible default for retrieval because it keeps context local while still pulling in dependencies.
• –PageRank-style scoring can help preserve structurally important memories, but it risks overpromoting central hubs unless edge types and decay are modeled carefully.
• –The shared vs private split is the most production-relevant idea here: multi-agent memory needs access control, not just similarity search.
• –Keeping the whole stack local with MCP and DuckDB is a strong privacy story, but it also means recall quality and incremental maintenance have to do all the heavy lifting.
• –This looks less like a finished answer to long-term memory and more like the right direction for reducing “fragmented forgetting” in practice.