Add memory as agent

When the LLM solves a task, it should summarize its learnings and call the `remember` tool. The remember tool should understand if this learning is something new and update the memory (vector database). A simple way to implement if a learning is new would be the distance of closest vector is > 0.2. When the LLM recieves a task, it queries the `memory` tool, to fetch the top k relevant memories, and uses them as context to solve the query.