Agent memory allows LLM agents to use earlier interactions when answering new queries. Existing methods often compress interaction histories into summaries or other LLM-generated representations. Repeated generation adds cost and can discard answer-bearing details before the system knows what a future query will require. We propose EdgeMem, an agent-memory method built around a simple principle: preserve original interaction turns and organize them through complementary content, temporal, and episodic cues. EdgeMem realizes this principle with a multi-anchor hypergraph constructed by lightweight local processing. Retrieval directly returns source evidence and reserves LLM use for final answer generation, combining structured access to multi-session histories with faithful retention of the original conversation. Experiments on LoCoMo and LongMemEval-S show strong retrieval and memory-grounded question answering; on LoCoMo, EdgeMem achieves the highest strict-judge score among seven reproduced systems under a shared prompt (61.01 versus 58.70), while construction and retrieval require no generative-LLM calls. Overall, EdgeMem shows that preserving and organizing source evidence provides an effective and efficient foundation for agent memory without generative memory management.
Figures & tables
Figure 1. (a) LLM-based summarization and retrieval can lose answer-bearing evidence, while original-evidence preservation keeps it directly retrievable. (b) Existing memory frameworks incur substantial generative-LLM token costs in both memory construction and per-query answering. Top: a side-by-side comparison in which LLM-based summarization and retrieval drop an answer-bearing detail before answering, while original-evidence preservation retrieves it directly. Bottom: a grouped horizontal bar chart on a logarithmic token axis; EdgeMem's construction cost is zero and its answering cost is the lowest.
Figure 2. EdgeMem overview. Top (construction): the interaction history is retained as source turns and processed by lightweight local annotation; time, co-occurrence, and session grouping derive the time, co-occurrence, and episode anchor keys of one multi-anchor hypergraph. Bottom (retrieval): query cues drive deterministic evidence channels over the Time, Co-occurrence, and Episode Sub-Hypergraphs. Their selected source turns are fused under a fixed budget and passed to a single reader, which returns an answer or an explicit refusal. The reader is the pipeline’s single LLM stage. A two-stage diagram. During memory construction, source turns are retained unchanged and locally annotated with entities, lexical content, episode membership, and dates. Time, shared-key, and session grouping induce the Time, Co-occurrence, and Episode Sub-Hypergraphs within one multi-anchor hypergraph. During retrieval, date, entity, and lexical query cues drive the three deterministic evidence channels. Their source turns are fused into one evidence pack, after which a single language-model reader produces an answer or refusal.
System
Build tokens ↓
Build calls ↓
Build time ↓
Tokens per question ↓
Calls per question ↓
Time per question (s) ↓
EdgeMem
0
0
1.8 min
1,000
1
1.3
CAM
1.27M
1,516
75 min
19,745
4.9
11.8
CompassMem
1.76M
1,500
43 min
41,348
21.7
45.8
MemGAS
0.46M
544
17 min
3,146
2
3.7
MemoryOS
4.49M
12,730
≥ 1.6 h
5,116
2
4.7
A-Mem
11.77M
17,768
10.1 h
5,857
2
5.2
Table 1. Comparison with released memory systems on LoCoMo, all numbers from our metered runs: (a) memory lifecycle cost, (b) answer quality under the shared answer prompt (Sections 4.1 and 4.2 ). In this and the following tables, the best value in each column is bold, the second best underlined, arrows mark the preferred direction, and n/a marks quantities the design does not define.
System
Judge (all questions) ↑
Exact refusal ↑
Lenient refusal ↑
Matched refusal ↑
EdgeMem
61.76
92.57
92.57
92.57
CAM
41.64
n/a
1.58
38.96
CompassMem
40.81
n/a
18.02
79.28
MemGAS
34.26
n/a
23.65
74.77
MemoryOS
31.37
n/a
12.61
75.00
A-Mem
25.19
n/a
2.93
84.68
Table 2. Behavior on the 446 LoCoMo adversarial questions under the three scoring arms of Section 4.1 .
Figure 3. (a) Memory quality (lenient judge, Table 1 b) against each released pipeline’s per-question token cost (Table 1 a); upper left is better. (b) The anchor ladder: the accuracy each anchor adds in deployment order (Table 3 ); diamonds show co-occurrence and time alone. Two small panels. Left: scatter of accuracy versus token cost on a logarithmic axis; EdgeMem sits in the upper left. Right: waterfall from a flat baseline through episode, co-occurrence, and time increments to the full chain, with diamonds for co-occurrence-only and time-only variants.
Configuration
SJ ↑
F1 ↑
BLEU-4 ↑
ROUGE-L ↑
BERTScore ↑
flat turn BM25
49.74
19.40
4.14
19.00
85.38
+ episode Sub-Hypergraph
58.18
21.94
4.88
21.41
85.80
+ co-occurrence Sub-Hypergraph
60.43
22.12
4.92
21.53
85.83
+ time Sub-Hypergraph
61.01
22.44
5.03
21.82
85.90
co-occurrence only
55.39
20.66
4.52
20.07
85.64
time only
50.19
19.54
4.19
19.16
85.40
Table 3. Anchor ladder on LoCoMo under the shared answer prompt: the deployed path adds one Sub-Hypergraph per rung (upper block); the lower block lists non-deployed single- and pair-anchor variants.
@3
@10
System
R ↑
N ↑
R ↑
N ↑
SJ ↑
Tokens per question ↓
Full History †
n/a
n/a
n/a
n/a
50.60
103,137
MPNet †
66.17
75.47
85.11
80.63
53.20
8,173
Contriever †
71.06
79.72
90.00
84.29
55.40
8,286
MPC †
60.00
70.90
80.00
76.59
53.80
8,457
RecurSum †
67.23
78.33
87.66
83.28
35.40
8,853
Table 4. LongMemEval-S: session-level Recall (R) and NDCG (N) at k∈{3,10} ( n=470 ), strict-judge accuracy (SJ), and reader input tokens per question.
Memory / summary
Retrieved evidence
Final answer
Q1: What gaming equipment did John buy or refurbish? gold: Sennheiser headphones, Logitech mouse, gaming desk
A-Mem
✓ notes faithfully retain headphones, mouse , and the refurbished desk
× partial retrieval: D23:10 omitted; mouse and desk never reach the reader
× “new devices and headphones”—misses mouse and desk
Q2: What kind of beer does McGee’s bar serve? gold: stout, lager
CAM
× key fact dropped: gist layer omits “McGee’s” and “lager” ; the exact leaf is pruned
× cannot recover what construction discarded
× “stout and light beers”—misses lager
EdgeMem
✓ original turns kept: D21:15 (great stout), D23:3 (“good lager at McGee’s”)
✓ D21:15, D23:3 packed
✓ “great stout and lager beer”
Table 5. Three audited failures, each traced from memory through retrieval to the answer, against EdgeMem ’s handling of the same questions: a faithful memory retrieved incompletely (A-Mem), a key fact destroyed during summarization (CAM), and a faithful store whose agentic navigation selects no evidence (CompassMem). ✓ : correct or preserved; × : missed or incorrect.