MemGPT
Agent memory architecture (Packer et al., 2024) that treats the model like an operating system and gives it tools to manage its own memory. The context window holds read-only system instructions, a writable working context and a first-in, first-out message queue. Messages pushed out of the queue go to recall storage, searchable by exact text, and long-term facts and documents go to archival storage, a vector database.
In michael-ryan's walkthrough (Stanford CS329Z, Oct 2026), the agent writes "birthday, boyfriend James" into working context and replaces it after a breakup, searches recall storage for "Six Flags", and pages through archival search results to answer a question. The model picks its own search terms; there is no hidden entity recognition. On deep memory retrieval it lifted GPT-4 from 32.1% to 92.5% and GPT-4 Turbo from 35.3% to 93.4%, and its accuracy held steady as more documents were retrieved. Ryan notes that rewriting the context invalidates the KV cache.
Ryan compares it with MemoryOS, which has short-, medium- and long-term tiers and a background system deciding what moves up; MemGPT leaves those decisions to the model. Sleep-time compute comes from the same line of work (sleep-time-compute).
Source: report