Memory OS of AI Agent

2026. 10. presenter: Joohyung Yun

PART 1

The Problem

Part 1 The Problem · long-term memory in multi-session conversation

Part 2 Existing Memory Methods and Their Failure Modes

Part 3 MemoryOS · the idea and the full procedure

Part 4 Experiments and Takeaways

01 / 25

Long-Term Memory in Multi-Session Conversation

fact recall · answer from what was said weeks ago, across sessions
personalization · act on what the user wants without being told again
one memory per user · spans separate sessions · topics grouped by the system, not by the user

02 / 25

Question Types in Long-Term Conversation Benchmarks

03 / 25

Fixed Context Window as the Root Cause

root cause · the model sees only what fits in the window, and earlier sessions do not

04 / 25

PART 2

Existing Memory Methods

Part 1 The Problem · long-term memory in multi-session conversation

Part 2 Existing Memory Methods and Their Failure Modes

Part 3 MemoryOS · the idea and the full procedure

Part 4 Experiments and Takeaways

05 / 25

Three Families of Memory Methods

four baselines in the experiments · each family focuses on one dimension of memory management

06 / 25

Failure Modes of Existing Memory Methods

07 / 25

PART 3

MemoryOS

Part 1 The Problem · long-term memory in multi-session conversation

Part 2 Existing Memory Methods and Their Failure Modes

Part 3 MemoryOS · the idea and the full procedure

Part 4 Experiments and Takeaways

08 / 25

Operating System Principles Reused in MemoryOS

same three principles · segments of pages · keep the hot, evict the cold · tiers by lifetime

09 / 25

Structure of MemoryOS

10 / 25

Running Example Used Throughout Part 3

one example, start to finish · a four-week conversation, traced through every part of the structure

11 / 25

MemoryOS on the Running Example

12 / 25

Three Storage Tiers and Their Contents

page {Q, R, T} → topic folder + summary + keywords → persona facts and traits

13 / 25

Dialogue Pages and the Dialogue Chain

chain · pages on one thread, linked or reset by the LLM · running summary kept as meta

14 / 25

Update From Short-Term to Mid-Term Memory

FIFO out of the 7-page queue · match score F > θ joins a topic folder, otherwise a new folder

15 / 25

Segment Heat, Eviction, and Promotion

promote · Heat > τ = 5 → the folder's pages and summary feed the long-term persona memory
evict · when MTM exceeds 2000 folders, the coldest one is deleted

16 / 25

Update to Long-Term Persona Memory

heat > τ · one LLM pass writes facts, 90-dim traits, and agent traits · queues of 100

17 / 25

Memory Retrieval and Response Generation

18 / 25

The Per-Turn Procedure of MemoryOS

LLM calls per turn · link judge · summaries · persona extraction · the answer

19 / 25

PART 4

Experiments

Part 1 The Problem · long-term memory in multi-session conversation

Part 2 Existing Memory Methods and Their Failure Modes

Part 3 MemoryOS · the idea and the full procedure

Part 4 Experiments and Takeaways

20 / 25

Experimental Setup

	GVD	LoCoMo
conversations	15 virtual users × 10 days · ≥ 2 topics per day · simulated user–assistant dialogues	≈ 300 turns · ≈ 9K tokens per conversation · questions in 4 types: single-hop · multi-hop · temporal · open-domain
metrics	retrieval accuracy (0 / 1) · correctness and coherence (0 / 0.5 / 1) · scored by DeepSeek-R1	F1 and BLEU-1 against the reference answer
backbones	GPT-4o-mini · Qwen2.5-7B	GPT-4o-mini · Qwen2.5-3B
baselines	TiM · MemoryBank · MemGPT · A-Mem (A-Mem* = re-implementation in the same environment; A-Mem = numbers reported by its authors)
MemoryOS settings	recent turns queue 7 pages · topic folders ≤ 2000 · User KB / Agent Traits queues 100 · promotion τ = 5 · folder match θ = 0.6 · heat weights α = β = γ = 1 · μ = 1e7 s retrieval: top-m = 5 folders · top-k = 5 (GVD) / 10 (LoCoMo) pages · top-10 persona entries
hardware	8 × H20 GPUs

21 / 25

Results on the LoCoMo Benchmark

22 / 25

Results on GVD and Efficiency

backbone	method	Acc.	Corr.	Cohe.
GPT-4o-mini	TiM	84.5	78.8	90.8
	MemoryBank	78.4	73.3	91.2
	MemGPT	87.9	83.2	89.6
	A-Mem	90.4	86.5	91.4
	MemoryOS	93.3	91.2	92.3
Qwen2.5-7B	TiM	82.2	73.2	85.5
	MemoryBank	76.3	70.3	82.7
	MemGPT	85.1	80.2	86.9
	A-Mem	87.2	79.5	87.8
	MemoryOS	91.8	82.3	90.5

GVD · retrieval accuracy / correctness / coherence, scored by DeepSeek-R1 · MemoryOS over A-Mem with GPT-4o-mini: +3.2% / +5.4% / +1.0%

23 / 25

Ablation over the Memory Modules

MTM matters most · LPM next · chain least · every tier removed costs something on every question type

24 / 25

Takeaways

Hierarchy, topic folders, heat

recent turns → topic folders → persona
FIFO moves a page down
match score picks its folder
heat promotes hot folders, evicts cold ones
four modules in one per-turn loop

Where the gains come from

LoCoMo avg F1 +49.11% over baselines
temporal, the hardest type: +118.8% F1
ablation: MTM > LPM > chain
the win is organization, not a bigger prompt

Cost and limits

4.9 LLM calls · 3.9K tokens per reply
A-Mem* 13 calls · MemGPT 17K tokens
hand-set sizes: STM 7 · 2000 folders · τ = 5
overlapping folders never merged
GVD is LLM-judged · A-Mem re-run lower

Discussion · heat rewards what was used often; when a user's interests shift, which signal should decide what to forget, and does a fixed 90-dimension trait schema carry over to other domains?

25 / 25