MemGPT: virtual memory for limited contexts¶

简体中文

Implements core, working, and archival state. Working-context pressure triggers an interrupt and page-out, with page-in on later tasks.

Paper information¶

Field	Details
Paper	MemGPT: Towards LLMs as Operating Systems
Organization	University of California, Berkeley
First public release	2023-10-12
Authors' code	The authors' public code, now Letta.
Adapter / method	`memgpt`
Local reproduction code	`src/auto_research/agent_research/`

Original paper¶

Motivation and method¶

Limited context windows forget information in long documents and conversations. MemGPT borrows operating-system virtual memory: resident core information, active working context, and an external archive. Function calls move data, and interrupts or heartbeats control continued reasoning and interaction with the user.

flowchart LR
    U["User or document stream"] --> W["Working context"]
    C["Core memory"] <--> W
    W --> I{"Context pressure interrupt"}
    I --> A["Archival memory"]
    A --> P["Retrieve and page in"]
    P --> W

Key figure from the paper¶

Figure 3: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ \text{if }|W_t|>B:\quad A_{t+1}=A_t\cup\operatorname{pageout}(W_t),\qquad W_{t+1}\leftarrow\operatorname{pagein}(q,A_{t+1}). \]

Paper results¶

Document analysis beyond the base-model window and multi-session conversations compare fixed-context and retrieval baselines, showing persistent memory and dynamic reflection. No production online A/B is reported.

Local reproduction¶

EvoMem mini, 120 episodes, seed 42, memory size 24. Working memory uses half the capacity; pressure triggers page-out and repeated tasks page in from archival memory.

Metric	Long-context baseline	MemGPT
Joint success	1.0000	1.0000
Average cost	64.5000	0.9200
Archival writes / page-ins / interrupts	0 / 0 / 0	108 / 96 / 108

auto-research agent-eval --method memgpt --benchmark evomem-mini \
  --episodes 120 --memory-size 24 --seed 42

See the stable metrics.

Reproduction boundaries¶

Implements memory tiers, page-in/out, and interrupt counts. Archival retrieval uses deterministic keys, not embedding services or LLM function calling. Cost is the shared proxy cost, not tokens or online latency.