Yurumi: the memory that learned to be an agent

From mem0 to agent

Yurumi started as an experiment: replacing mem0 (expensive and slow) with something local, fast, and within budget. A Qdrant store with triple embedding and simple ReAct retrieval. It worked. But I wanted more.

The big leap: F1 to F4

In less than two weeks, Yurumi stopped being passive memory and became a complete agentic system:

F1 — Intent and Actions: Each memory is classified into 7 types. The agent itself executes real actions — notifying on Telegram, running commands, posting memory. All wired through INGEST→DECIDE→ACT→FEEDBACK→LEARN.

F2 — MCP and REST: Every action gained MCP tools and REST endpoints. I can create/release holds, tick the agent, even restart WSL remotely (with approval).

F3 — Learning by itself: consolidate dry-run groups latent duplicates without applying, expire cleans old holds, and the 5-min cron only wakes up when there’s news.

F4 — Identity: Each machine/agent has a unique ID in the AgentRegistry. The system knows who everyone is and what they can do.

Graphify: the graph that became a brain

A knowledge graph connecting real entities with evidenced edges. It finds god nodes (the hubs connecting everything), surprising connections between distant communities, suggested questions crossing projects, and an interactive HTML visualization.

Vectorize: the fallback that saved the night

On the dawn of 30/08, Qdrant went down (WSL out of RAM). Yurumi had a fallback: Cloudflare Vectorize. Insert worked, query worked, nobody noticed.

What I learned

Memory isn’t just about storing — it’s deciding what to do with what was stored. A passive store is a database. An agentic memory system is a partner that remembers, acts, and learns.

The F1-F4 architecture is portable. What works for memory works for any agent. I’m taking this pattern to all projects.