Agents in a society can reuse each other's experience, but indiscriminate memory use increases processing cost and can reduce task success. We present SEDM (Selective Self-Evolving Distributed Memory), a framework that selects both what agents share and when recipients use it. On the sender side, SEDM admits memories through paired execution tests, ranks them by relevance and measured utility, and abstracts them for cross-domain reuse. On the recipient side, Calibrated Memory Gating (CMG) predicts success with and without memory in a single forward pass, learning from paired outcomes including ties. It injects memory only when the calibrated success gain outweighs a context-cost penalty. SEDM exceeds the strongest evaluated baseline on LoCoMo open-domain questions by 15.1 F1 points and outperforms G-Memory on FEVER and HotpotQA with 16-32% fewer prompt tokens. In a controlled HotpotQA study, CMG achieves success comparable to always using memory with 18% fewer agent input tokens. Within the full SEDM pipeline on LoCoMo, CMG outperforms both always-use and no-memory policies for two SEDM memories while using 19-31% fewer input tokens than always-use. These results show that selective use can recover value from shared memories that hurt performance when injected indiscriminately.
Figures & tables
Figure 1: Memory strategies for agent societies. (a) Without memory, agents do not reuse experience. (b) With a fixed memory, all past experiences are stored and retrieved memory is always injected. (c) SEDM is selective on the sender side, through verifiable admission (write) and utility-guided organization, and on the recipient side, where the CMG gate decides whether the receiving agent uses the retrieved memories M(q) .
Figure 2: Full SEDM architecture. SEDM integrates SCEC-based verifiable write admission, utility-aware memory scheduling, cross-domain knowledge diffusion, and calibrated memory gating (CMG), which decides whether to USE or SKIP retrieved memories.
Method
Single Hop
Multi-Hop
Open Domain
Temporal
Adversarial
F1 ↑
B1 ↑
F1 ↑
B1 ↑
F1 ↑
B1 ↑
F1 ↑
B1 ↑
F1 ↑
B1 ↑
LoCoMo ( Maharana et al., 2024 )
40.4
29.1
25.0
19.8
12.0
11.2
18.4
14.8
69.2
68.8
ReadAgent ( Lee et al., 2024 )
9.7
7.7
9.2
6.5
5.3
5.1
12.6
8.9
9.8
9.0
MemoryBank ( Zhong et al., 2024 )
6.6
5.2
5.0
4.8
5.6
5.9
9.7
7.0
7.4
6.5
MemGPT ( Packer et al., 2024 )
41.0
34.3
26.7
17.7
9.2
7.4
25.5
19.4
43.2
42.7
A-Mem ( Xu et al., 2025 )
44.7
37.1
27.0
20.1
12.1
12.0
45.9
36.7
50.0
49.5
Table 1: Results on LoCoMo. F1: token F1; B1: BLEU-1. Best result per column in bold. SEDM here injects every retrieved memory (no CMG); CMG is evaluated in Sections 4.5 – 4.6 .
FEVER
HotpotQA
Method
Acc.
Prompt
Compl.
EM
Prompt
Compl.
No memory
57
1.65M
24K
34
2.46M
29K
G-Memory
62
3.62M
109K
38
4.63M
114K
SEDM w/o CMG, adding components
+ SCEC admission
64
2.19M
53K
37
3.52M
52K
+ self-scheduling
66
2.47M
53K
39
3.88M
55K
Table 2: Accuracy and token cost on FEVER and HotpotQA. Prompt/Compl.: total prompt and completion tokens. SEDM rows add components cumulatively. SEDM here injects every retrieved memory (no CMG); CMG is evaluated in Sections 4.5 – 4.6 .
Collect ↓ / Test →
FEVER
HotpotQA
LoCoMo
No memory
57
34
–
FEVER
66
41
38.1
HotpotQA
64
39
38.6
LoCoMo
65
34
37.6
Table 3: Cross-domain evaluation of SEDM w/o CMG. Rows give the dataset on which memory is collected, and columns the dataset on which it is tested (FEVER: accuracy; HotpotQA: EM; LoCoMo: overall token F1). The LoCoMo scores come from the original cross-domain runs and are not directly comparable to the per-type scores in Table 1 .
Gate
Success [95% CI]
Injection rate
Agent input tokens
Gate latency (s)
Always skip (no memory)
0.249 [0.213, 0.286]
0%
1,629
–
Always use
0.333 [0.292, 0.376]
100%
3,736
–
Similarity threshold ( τ = 0.2)
0.333 [0.292, 0.376]
100%
3,736
–
Single-token SFT (winner labels)
0.333 [0.292, 0.376]
100%
3,736
0.24
Winner head (winner labels)
0.333 [0.292, 0.376]
100%
3,736
0.12
CoT gate (generative)
0.262 [0.225, 0.301]
9%
1,734
3.09
Table 4: Gate comparison in the controlled HotpotQA study (test split, 489 questions, seed 42). Token savings are relative to always use. All learned gates share the Qwen3-4B backbone; latency was measured on a shared A40, and the two single-pass heads differ only by server load.
Figure 3: Quality–cost trade-off of CMG. Each blue curve sweeps the price λ from large (left, memory almost never used) to λ=0 (right). The dashed line is the expected result of random gating between no memory and always using memory; black squares mark always using memory. (a) HotpotQA test split, seed 42, with the generative CoT gate for reference. (b, c) LoCoMo test conversation, mean of three seeds, for the two SEDM memories. In all three settings the CMG curve lies above random gating.
Setting
Succ.
F1
Inj.
Tok.
SEDM memory: admission only
No memory
0.544
0.556
0%
1,476
SEDM, no gate
0.539
0.564
100%
2,790
SEDM + CMG (val. λ=0 )
0.556
0.577
65%
2,263
SEDM + CMG ( λ=0.05 )
0.568
0.579
30%
1,825
Oracle
0.606
–
–
–
Table 5: CMG inside the full SEDM pipeline on LoCoMo (test conversation 2, 193 questions). CMG rows are means over three seeds. Succ.: SBERT ≥ 0.7; F1: token F1; Inj.: injection rate; Tok.: mean agent input tokens per question. The two “No memory” rows are independent runs of the same agent.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Head
NLL
Brier
ECE
p\textscskip
0.543 → 0.538
0.177 → 0.177
0.051 → 0.049
p\textscuse
0.584 → 0.579
0.198 → 0.196
0.060 → 0.044
Appendix
Table 6: Top: test calibration of CMG before → after temperature scaling (seed 42). Bottom: fitted temperatures and test success rate across seeds. For seed 43, validation selected λ=0.05 .
Figure 4: Injection rate of CMG as the price λ grows (test splits, mean of three seeds). The legend gives the effect of always using memory on the test split. The more a memory helps, the more often CMG injects it at every price.
Split
n
S\textscskip
S\textscuse
00
01
10 / 11
Train
2,962
0.249
0.307
1,814
409
239 / 500
Validation
496
0.244
0.288
312
63
41 / 80
Calibration
496
0.254
0.310
300
70
42 / 84
Test
489
0.249
0.333
294
73
32 / 90
Appendix
Table 7: Paired outcomes per split. 01: only use succeeds; 10: only skip succeeds.
λ
Seed 42
Seed 43
Seed 44
0
0.329 / 69%
0.323 / 82%
0.325 / 77%
0.05
0.315 / 40%
0.301 / 45%
0.317 / 56%
0.1
0.288 / 24%
0.282 / 19%
0.305 / 38%
0.2
0.282 / 13%
0.270 / 5%
0.290 / 19%
0.5
0.262 / 3%
0.249 / 0%
0.260 / 2%
1
0.249 / 0%
0.249 / 0%
0.249 / 0%
Appendix
Table 8: Test success / injection rate of calibrated CMG for each λ and seed. The shape of the curve is stable; the injection rate at a given λ varies across seeds, so gates are best compared at matched injection rates.
Gate
Success
Injection
Latency (s)
Always skip
0.275
0%
–
Always use
0.245
100%
–
Winner head
0.270
–
–
Single-token SFT
0.270
–
–
CoT gate
0.270
5%
2.70
CMG, λ=0
0.275
1%
0.06
Appendix
Table 9: Pilot results (200 test questions, seed 42). CMG seeds 42/43/44: 0.275 / 0.275 / 0.255.
Self-evolving multi-agent systems (MAS) have emerged as a promising route to LLM agents that continually improve from experience, with persistent memory at their foundation. However, existing designs almost exclusively adopt a centralized repository shared across agents, incurring communication and coordination overhead, raising privacy concerns, and collapsing agent diversity. We propose DecentMem, a decentralized memory framework in which each agent maintains its own dual-pool memory -- an exploitation pool of consolidated past trajectories and an exploration pool of LLM-generated candidates for unseen contexts. The two pools are reweighted online based on stage-wise feedback from an LLM-as-a-judge. Theoretically, we prove that this design guarantees global reachability of the solution space and achieves O(logT) cumulative regret, matching the stochastic bandit lower bound up to constants. In practice, across three MAS frameworks (AutoGen, DyLAN, AgentNet), three Qwen3 backbones (4B/8B/14B), two Gemma4 backbones (E2B/E4B) and five benchmarks spanning math, code, QA, and embodied tasks, DecentMem improves average accuracy by up to 23.8% over the strongest centralized memory baseline and by up to 52.5% over the no-memory baseline, while reducing token usage by up to 49%.
Agent memory faces heterogeneous access needs: a single-hop question may require one piece of evidence, whereas a multi-hop question must combine evidence from multiple sources. Predefined memory workflows cannot adapt to these varying needs. Recent adaptive methods search or learn over memory components and their compositions, but the design space itself remains predefined. We introduce MemCodex, a self-evolving hierarchical memory system that organizes experience into executable memory programs for summaries, relational knowledge, reusable skills, and latent memory. Open-ended program evolution searches the open design space of layer programs by rewriting how each layer is constructed, indexed, retrieved, and routed, thereby adapting both within-layer implementations and cross-layer composition. At query time, reads traverse the hierarchy from coarse to fine and stop once sufficient evidence is found, descending to the original history when needed. We further develop MemArena, a unified runtime that places heterogeneous data and memory systems behind a common interface. MemCodex improves average task success by 10.1% relative to the strongest adaptive-memory baseline, while using 3.4x fewer context tokens and achieving 2.1x faster inference.
Xiaoqiang Wang, Bang Liu
Université de Montréal · Mila – Quebec AI Institute
Recent memory agents improve LLMs by extracting experiences and conversation history into an external storage. This enables low-overhead context assembly and online memory update without expensive LLM training. However, existing solutions remain passive and reactive; memory growth is bounded by information that happens to be available, while memory agents seldom seek external inputs in uncertainties. We propose autonomous memory agents that actively acquire, validate, and curate knowledge at a minimum cost. U-Mem materializes this idea via (i) a cost-aware knowledge-extraction cascade that escalates from cheap self/teacher signals to tool-verified research and, only when needed, expert feedback, and (ii) semantic-aware Thompson sampling to balance exploration and exploitation over memories and mitigate cold-start bias. On both verifiable and non-verifiable benchmarks, U-Mem consistently beats prior memory baselines and can surpass RL-based optimization, improving HotpotQA (Qwen2.5-7B) by 14.6 points and AIME25 (Gemini-2.5-flash) by 7.33 points.