ECHO: A Cognitively Inspired, Auditable Memory Plane for Long-Horizon Agents

Yu Qian; Hong Miao; Boyang Guo; Tingyi Jiang; Shan Zhao; Tianxing Le; Lintian Li; Meng Liu

cs.AIAug 22, 2026

ECHO: A Cognitively Inspired, Auditable Memory Plane for Long-Horizon Agents

Authors: Yu Qian, Hong Miao, Boyang Guo, Tingyi Jiang, Shan Zhao, Tianxing Le, Lintian Li, Meng Liu

Organizations: XDream Robotics, Building T2, Moli Community, No. 1, Lane 188, Yuren Road, Pudong New Area, Shanghai 200120, China

Abstract

Long-horizon agents need memory that identifies relevant experience, resolves revisions, and exposes checkable provenance. We present ECHO (Embodied Context and History Orchestration), an auditable memory architecture and service prototype inspired by episodic encoding, consolidation, contextual reinstatement, reconsolidation, and executive control. This is functional inspiration, not neural equivalence; the empirical analysis focuses on retrieval and context construction. Development runs reach 96.29% Hit@10 and 73.64% turn Recall@5 on 1,536 LoCoMo category 1-4 questions, and 97.60% Hit@10, 88.84% turn Recall@5, and 88.71% session Recall@5 on all 500 LongMemEval-S questions. A five-history BEAM gate fails, and in a separate matched 91-question QA sample Mem0 OSS scores 64.84% versus ECHO's 41.76% (exact McNemar p = 0.00107), with a history-cluster interval crossing zero. A post-hoc audit found source-specific phrases in the query-expansion rules. Although no gold answer field entered the runtime, expansion-enabled retrieval scores are therefore descriptive development measurements, not independent confirmation.

Figures & tables

Cognitive principle	Engineering commitment
Episodic encoding	Immutable, source-ordered experience traces
Semantic consolidation	Typed state projected separately from raw episodes
Reconsolidation	Append revisions; never erase the prior trace
Contextual recall	Cue-driven candidates resolved in valid/known time
Executive control	Operation routing, selective expression, and abstention

TABLE I: Functional cognitive inspiration and testable system commitments.

Panel of Figure 2 — Fig. 1: ECHO at a glance. The upper lane translates a functional cognitive loop into the auditable memory plane below; colors distinguish experience, projected state, temporal authority, discovery, provenance, and realization. Solid arrows are runtime flow and dashed arrows are testable correspondences. The feedback loop admits only new experience or authorized correction, never an unverified retrieval result.

Panel of Figure 3 — Fig. 2: Target ECHO runtime architecture and current integration boundary. Solid arrows denote the intended synchronous authority path; dashed arrows denote asynchronous projection, observability, and authorized updates. The empirical analysis focuses on retrieval and context construction. Cognitive labels are functional correspondences, not claims of anatomical or neural equivalence.

Operation	Internal requirement	Final surface
Fact	Minimal sufficient support	Minimal supported answer
Count	Complete member set, deduplicated by identity, and verified count	Numeric count; unit only if needed
List	Complete deduplicated set; requested order retained	Requested items only
Temporal arithmetic	Both endpoints, normalized unit, verified calculation	Requested duration or date only
Temporal lookup	Valid/known-time view and supported target state	Time-scoped fact only

TABLE II: Internal and visible realization contracts.

Dataset	Reg./eval.	Hist.	Scope
LoCoMo	1,986/1,982	10	categories 1–4 primary; category 5 separate
LongMemEval-S	500/500	500	all six released question types
BEAM development	20/18	1	one 100K row; pilot only
BEAM fresh gate	91/89	5	preregistered histories; retrieval only

TABLE III: Frozen ECHO retrieval scopes. Registered/evaluated counts are both shown to prevent denominator drift.

Dataset scope	Eval.	Hit@10 (%)	Turn R.@5 (%)	Turn R.@10 (%)	Sess. R.@5 (%)	Sess. R.@10 (%)
LoCoMo cat. 1–4	1,536	96.29	73.64	84.48	–	–
LongMemEval-S	500	97.60	88.84	91.74	88.71	93.46

TABLE IV: Cutoff-aligned evidence retrieval, not QA accuracy. Hit@10 is question-level; turn/session R.@5 and R.@10 are micro annotated-unit coverage recomputed from the same frozen Top-10 rows.

Panel of Figure 7 — Fig. 3: Question-level hits conceal incomplete evidence coverage. The gap is largest on the local BEAM pilot. LoCoMo uses the category-1–4 primary slice.

History row	Evaluated	Hit@10 (%)	Turn R. (%)	Sess. R. (%)	MRR
6	18	94.44	47.06	51.11	.6308
16	18	94.44	58.18	67.50	.7102
12	18	61.11	22.99	40.00	.4611
9	17	82.35	44.44	58.97	.6123
5	18	61.11	42.62	46.67	.4894
Aggregate	89	78.65	40.80	52.34	.5804

TABLE V: Preregistered fresh-history BEAM generalization for ECHO Top-10 retrieval. The prespecified gate required at least 80 evaluated questions, zero projection failures, Hit@10

\geq 85\%

, turn recall

\geq 50\%

, and session recall

\geq 60\%

Sample	$n$	ECHO correct (%)	Mem0 correct (%)	$\Delta_{\mathrm{ECHO-Mem0}}$ (pp)	exact $p$
LoCoMo	65	29 (44.62; 33.17–56.66)	46 (70.77; 58.80–80.42)	-26.15	.00333
LongMemEval	6	3 (50.00; 18.76–81.24)	2 (33.33; 9.68–70.00)	+16.67	1.00000
BEAM local 100K	20	6 (30.00; 14.55–51.90)	11 (55.00; 34.21–74.18)	-25.00	.12500
All sampled questions	91	38 (41.76; 32.16–52.02)	59 (64.84; 54.61–73.86)	-23.08	.00107

TABLE VI: Matched sampled end-to-end QA. Both arms use the same binary Qwen answerer/judge protocol. Parentheses contain Wilson 95% intervals. Mem0 provenance retrieval metrics remain unobservable and are not assigned zero.

Source/system	Dataset	Reported metric	Score	Protocol distinction
Mem0 managed platform	LoCoMo	platform score	92.5	Top-200, production stack with proprietary optimizations
Mem0 managed platform	LongMemEval	platform score	94.4	same managed stack; not OSS and not ECHO retrieval recall
Mem0 managed platform	BEAM-1M / 10M	platform score	64.1 / 48.6	official larger scales, not the local one-row 100K pilot
Mem0 paper, base	LoCoMo	overall LLM-judge	$66.88\pm 0.15$	GPT-4o-mini-era paper protocol and averaged judge runs
Mem0 paper, graph	LoCoMo	overall LLM-judge	$68.44\pm 0.17$	graph memory; different storage, reader, judge, and cutoff
Zep / full context	LoCoMo	overall LLM-judge	$65.99/72.90$	author-reported baselines in the Mem0 paper

TABLE VII: Primary-source Mem0 numbers shown for orientation only. None is metric-compatible with ECHO retrieval in Table IV or the matched sampled QA in Table VI .

Variant	Hit@10 (%)	Turn R. (%)	Sess. R. (%)	MRR	Mean query (ms)	Mean ctx. (KiB)
Full v11	100.00	66.04	78.95	0.811	927.8	30.6
Dense-only	100.00	56.60	65.79	0.585	796.4	28.3
Lexical-only	83.33	50.94	78.95	0.463	825.6	36.2
No query expansion	83.33	39.62	60.53	0.644	427.5	30.5
No session-support weight	100.00	62.26	78.95	0.784	994.3	30.5

TABLE VIII: Load-bearing BEAM-100K retrieval ablations (

n=18

evaluated questions). Dense-only disables lexical and entity routes; lexical-only disables dense and entity routes. Latency includes local query embedding. Results are a paired local diagnostic, not a BEAM-1M/10M claim.

Panel of Figure 12 — Fig. 4: Normalized worker throughput on the fixed Windows/WSL2 Docker host. Stateful TopicBucket saturates near four workers; pure C++ ROSE reranking continues scaling. These are worker-path measurements, not QA accuracy.

Agent Memory Management

Agent memory management studies how AI agents store, update, retrieve, compress, and forget information across long-running tasks.

See all 443 papers

Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation
Oct 13, 2026Yufeng Li, Sheng Yao, Peng Zeng +5Agent Memory Management Long-Term Conversational Memory
East China Normal University, Shanghai, China · University of Southern California, Los Angeles, USA · Zhejiang University, Hangzhou, China +1
Eywa: Provenance-Grounded Long-Term Memory for AI Agents
May 29, 2026Resham JoshiLLM Agent Memory Persistent Memory for Language Models
Eywa
EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents
Jun 4, 2026Yilong Li, Suman Banerjee, Tong ChePersistent Memory for Language Models Long-Term Conversational Memory
University of Wisconsin–Madison · NVIDIA Research

Oct 13, 2026J/K move · Enter open · S save

Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation

Yufeng Li, Sheng Yao, Peng Zeng, Dezhang Kong+4

East China Normal University, Shanghai, China · University of Southern California, Los Angeles, USA · Zhejiang University, Hangzhou, China+1

Agent Memory Management Long-Term Conversational Memory Memory Agent Benchmarks

Open paper AskPDF arXiv

Long-term memory supports agents across sessions. Repeated retrieval and write-back can let one historical claim occupy multiple retrieval positions and influence current answers without independent evidence. Existing work examines retrieval, updating, and question answering, leaving same-source record accumulation and post-correction persistence underexplored. Motivated by research in social cognition and memory psychology on repeated exposure, retrieval, and revision, we define Memory Echo as the excess retrieval exposure gained by historical information from the same source without additional independent evidence, and introduce MEchoBench to evaluate its emergence, correction resistance, persistence, and effect on answers. To address this phenomenon, we introduce the Family--Provenance--Authority framework (FPA). Specifically, FPA selects one representative per source family, conditions the eligibility of old records for the answer context on verified corrections, and constrains final answers with authoritative evidence. In paired replay on 190 fixed questions across four memory frameworks, FPA raises accuracy from 0.825 to 0.920 with DeepSeek V4 Flash and from 0.782 to 0.916 with Kimi K2.6; mean Decision Contamination Effect falls from 0.106 to 0.030 and from 0.109 to 0.038, respectively. These results show that agent memory can preserve traceable history without allowing repeated historical information to govern current decisions. The code is available at https://github.com/OrcasSam/memory-echo-benchmark.

Abstract

Agent Memory Management

Explore similar work

Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation

Eywa: Provenance-Grounded Long-Term Memory for AI Agents

EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents