---
title: "MemGPT: Towards LLMs as Operating Systems"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: "paper is CC BY 4.0 (verified on arXiv abs page 2026-10-18; icon by-4.0.png, link text view license, http://creativecommons.org/licenses/by/4.0/); note is our own summary"
source: https://arxiv.org/abs/2310.08560
source_license: "CC BY 4.0"
attribution: "Summary by iinventedlight (AI-assisted). Source: Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez, \"MemGPT: Towards LLMs as Operating Systems\" (CC BY 4.0)."
condensed_by: "iinventedlight (Grok Bot, AI-assisted)"
fact_checked: 2026-10-18
date: 2026-10-18
status: published
category: arxiv
arxiv_id: 2310.08560
alias: 2310.08560-memgpt
checked_on: 2026-10-18
sources:
  - id: S1
    title: "MemGPT: Towards LLMs as Operating Systems"
    authors: "Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez"
    url: https://arxiv.org/abs/2310.08560
    version_read: v2 (12 Feb 2024)
    accessed: 2026-10-18
    license: CC BY 4.0
    attribution: "Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez, MemGPT: Towards LLMs as Operating Systems, arXiv 2023, CC BY 4.0, https://arxiv.org/abs/2310.08560"
---
# MemGPT: Towards LLMs as Operating Systems

## Summary
Packer, Wooders, Lin, Fang, Patil, Stoica, and Gonzalez start from a fixed context window, which blocks long conversations and long documents. MemGPT borrows the operating-system idea of a memory hierarchy: move data between a small fast tier and a larger slow tier so the model can work past the window it can see at once. The system decides what to bring into the window, and it uses interrupts to pass control back and forth with the user. The abstract evaluates that design in two places where a short window is a hard limit. One is document analysis, on documents much longer than the window. The other is multi-session chat, where the agent is meant to remember, reflect, and change across many conversations with the same person. The abs page lists no conference. This note follows that abstract page only. The abstract states no numeric score.

## When to apply
- The question is how a published system pages a long working memory in and out so an assistant can keep a document or a conversation that does not fit in one window.
- The question is which product setting turns memory on. This paper describes a research system, not a switch in an app.

## Costs and limits
- The abstract says MemGPT can analyze documents beyond the window and can carry a chat across sessions. It does not, on this page, report an accuracy or a memory-size limit. This note does not invent one.
- Comments say code and data are available at a URL. That URL's license was not checked, so it is not a source here. The PDF was not re-read.

## Evidence (detail)
- The abs page (v2, 12 Feb 2024) proposes virtual context management, moving data between fast and slow memory so a limited window can stand in for a larger store (Abstract).
- MemGPT, as the abstract defines it, manages those tiers and uses interrupts for control flow between the system and the user (Abstract).
- Evaluation domains named on the abstract: document analysis beyond the context window, and multi-session chat in which agents remember, reflect, and evolve (Abstract).
- Dateline: submitted on 12 Oct 2023 (v1), last revised 12 Feb 2024 (this version, v2). No journal reference on the page.

## Facts
- id: F1
  claim: MemGPT manages a hierarchy of memory tiers, paging information in and out of a fixed context window, and uses interrupts to pass control between the system and the user.
  confidence: medium
  date: 2026-10-18
  sources: [S1 Abstract]
  license: CC BY 4.0
- id: F2
  claim: The abstract evaluates that design on document analysis beyond the context window and on multi-session chat where an agent is meant to remember, reflect, and change across sessions. The abstract states no numeric score.
  confidence: medium
  date: 2026-10-18
  sources: [S1 Abstract]
  license: CC BY 4.0
- id: F3
  claim: The arXiv abs page shows CC BY 4.0. The dateline says submitted on 12 Oct 2023 (v1) and last revised on 12 Feb 2024 (v2), with no journal reference.
  confidence: high
  date: 2026-10-18
  sources: [S1 abs page]
  license: CC BY 4.0

## Related
- The library miss is how to try making an assistant that can keep a project. This note covers paging context in and out. It does not replace a lookup when the fact lives outside the stored memory.

## Review notes
- Staging draft. Paraphrase of the arXiv abstract page checked 2026-10-18. The PDF was not re-read.
- License on the abs page: icon by-4.0.png, link text "view license", http://creativecommons.org/licenses/by/4.0/.
- The credit line keeps the venue as arXiv 2023, the year of the first posting. version_read is v2 (12 Feb 2024).