YourMemory: MCP Memory That Forgets on Purpose
YourMemory is an MCP server that shipped on Product Hunt today with one unusual pitch. It is designed to forget. The author modeled the retention curve on the Ebbinghaus forgetting curve from cognitive science, so every piece of agent memory decays unless it gets reinforced, and only dies when all its graph connections fall below a pruning threshold. The claim. 84 percent token reduction and 52 percent recall on LoCoMo, which is roughly double what most commercial agent memory services manage.
Most agent memory stacks today are built on the assumption that memory is permanent and vector search will sort the relevance later. That is exactly why every serious agent developer eventually hits context rot. The model spends tokens reading things that looked relevant but are not, because vector similarity does not know what the task actually needs. YourMemory tries the opposite. Be aggressive about forgetting. Keep a graph of which memories connect to which. Let importance scores slow decay for the critical architectural decisions. Delete the rest.
The implementation choices are good. Local-first with DuckDB as the store, which means no vector DB vendor lock-in and no data ever leaves the machine. MCP-native, so Claude Code, Codex, Cursor, and anything else that speaks MCP can use it without a wrapper. Graph engine at v1.3.0 that identifies related architectural concepts beyond keyword matching. Open source and free.
The deeper point. Agent memory is becoming the hardest research problem in the field after reasoning itself, because every memory system has to make a bet about what the agent will need next. YourMemory bets on biological decay. Mem0 bets on structured profiles. Zep bets on temporal graphs. If you are building an agent that runs longer than a week of conversations, all three are worth trying. Start with the one that forgets on purpose.
Repo https://github.com/sachitrafa/YourMemory
← Back to all articles
Most agent memory stacks today are built on the assumption that memory is permanent and vector search will sort the relevance later. That is exactly why every serious agent developer eventually hits context rot. The model spends tokens reading things that looked relevant but are not, because vector similarity does not know what the task actually needs. YourMemory tries the opposite. Be aggressive about forgetting. Keep a graph of which memories connect to which. Let importance scores slow decay for the critical architectural decisions. Delete the rest.
The implementation choices are good. Local-first with DuckDB as the store, which means no vector DB vendor lock-in and no data ever leaves the machine. MCP-native, so Claude Code, Codex, Cursor, and anything else that speaks MCP can use it without a wrapper. Graph engine at v1.3.0 that identifies related architectural concepts beyond keyword matching. Open source and free.
The deeper point. Agent memory is becoming the hardest research problem in the field after reasoning itself, because every memory system has to make a bet about what the agent will need next. YourMemory bets on biological decay. Mem0 bets on structured profiles. Zep bets on temporal graphs. If you are building an agent that runs longer than a week of conversations, all three are worth trying. Start with the one that forgets on purpose.
Repo https://github.com/sachitrafa/YourMemory
Comments