MemGPT's core insight was borrowed from operating systems: treat the context window like RAM and external storage like disk, and let the LLM page memories in and out itself. The research became Letta, and the idea became mainstream — most serious agents now write and retrieve their own memories. The question is how much machinery you want to run for it.
Keeping the idea, dropping the ops
- Self-managed memory: with Kit for AI's MCP tools, the agent itself decides to remember a fact mid-conversation and recall it later — the MemGPT behaviour, no framework.
- Paging beyond the window: recall retrieves only the relevant memories into context, ranked by hybrid search, instead of refilling the window wholesale.
- Durability: memories are versioned, deduplicated records in the cloud — not state inside one runtime.
What you don't have to run
No agent server, no vector DB, no embedder to host. One config block gives an MCP agent its memory tools; one REST call does the same for your code. The document side — the files and pages your agent should know — lands in the same knowledge bases with the same retrieval.
Which should you pick?
To study or extend the memory-management architecture itself, MemGPT's lineage (Letta) is where that work lives. To give real agents durable memory today with the least moving parts, Kit for AI is the shortcut. Free to start.