Memory Papers
Memory Architecture

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

AuthorsYu Qian, Hong Miao, Boyang Guo et al.

arXiv 20262026
arXivPDF

TL;DR

ECHO uses a bitemporal ledger plus provenance-closed evidence packing to deliver 97.60% Hit@10 and 88.84% turn Recall@5 on LongMemEval-S.

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Headline scores hide missing evidence and stale answers

Long-horizon agents often rely on Hit@k metrics that “hide how much annotated evidence was recovered and whether a memory API exposes the provenance needed to measure that coverage.”

This means a system can score 100% Hit@10 yet miss multi-turn support, let a semantically close stale value outrank the current revision, or answer incorrectly even when some relevant evidence is retrieved.

HOW IT WORKS

ECHO: Embodied Context and History Orchestration

ECHO’s core mechanism wires Immutable events, Typed projection, Bitemporal ledger, Candidate discovery, Provenance closure, and Derive then realize into a continual loop of encoding, organization, recall, and evolution.

Think of ECHO as a governed card catalog plus ledger: events are the raw cards, the bitemporal ledger is the timestamped catalog, and provenance-closed packs are carefully assembled folders handed to the reader.

This design lets ECHO decide currentness and conflicts via the ledger, not similarity, and ensures every answer is tied to auditable provenance and a context hash rather than a plain context window dump.

HOW IT WORKS

ECHO Architecture Flow

ECHO’s architecture starts from Immutable events and Typed projection, feeding a Bitemporal ledger that encodes valid and transaction time with revision status.

At query time, Candidate discovery proposes lexical, dense, and typed routes, but the Bitemporal ledger alone decides currentness and conflicts before Provenance closure packs atomic evidence sets.

Finally, Derive then realize computes internal results and emits only the contracted surface answer or abstention, keeping evidence completeness and answer verbosity explicitly separated.

DIAGRAM

ECHO Query-Time Memory Retrieval Pipeline

This diagram shows how ECHO processes a public question through operation planning, multi-route discovery, ledger resolution, provenance-closed packing, and selective realization.

DIAGRAM

ECHO Evaluation and Gate Protocol

This diagram shows how ECHO’s evaluation pipeline separates frozen development runs, fresh BEAM gates, and matched Mem0 QA under explicit metric and leakage checks.

PROCESS

How ECHO Handles a Long-Horizon Question

  1. 01

    Immutable events

    ECHO first records interaction history as Immutable events with source order, actors, and timestamps, ensuring episodes are never overwritten.

  2. 02

    Typed projection

    Projection workers convert raw turns into typed propositions and state revisions, attaching provenance links back to the originating events.

  3. 03

    Bitemporal ledger

    The Bitemporal ledger appends revisions with valid time, transaction time, status, and generation, enforcing that semantic similarity cannot determine factual currentness.

  4. 04

    Candidate discovery and provenance closure

    At query time, Candidate discovery unions lexical, semantic, and typed routes, then Provenance closure packs atomic evidence sets under token budgets for the reader.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Typed bitemporal authority component

    ECHO specifies and unit-tests a Bitemporal ledger where valid time, transaction time, revision state, provenance, and conflicts are explicit, and similarity cannot decide currentness.

  • 02

    Worker-oriented retrieval pipeline

    ECHO implements a worker-oriented Candidate discovery pipeline and evaluates retrieval, Provenance closure, and bounded-context behavior at explicit system boundaries.

  • 03

    Protocol-matched Mem0 comparison

    ECHO provides a 91-question protocol-matched QA comparison where Mem0 OSS scores 64.84% versus ECHO’s 41.76%, with exact McNemar p = .00107 and history-cluster analysis.

RESULTS

By the Numbers

Hit@10

97.60%

LoCoMo Hit@10 is 96.29%, so LongMemEval-S is +1.31 percentage points higher for ECHO.

Turn Recall@5

88.84%

On LoCoMo categories 1–4, ECHO’s turn Recall@5 is 73.64%, a 15.20 point gap versus LongMemEval-S.

Turn Recall@10

91.74%

LongMemEval-S session Recall@5 is 88.71%, showing strong multi-session coverage alongside turn Recall@10.

QA accuracy

41.76%

Mem0 OSS reaches 64.84% on the same 91 questions, a −23.08 percentage point difference for ECHO.

These numbers come from frozen ECHO retrieval on LoCoMo categories 1–4 and LongMemEval-S, plus a matched 91-question QA sample against Mem0 OSS. They show that ECHO’s memory plane delivers high evidence coverage while its current reader underperforms Mem0 on strict QA accuracy.

BENCHMARK

By the Numbers

These numbers come from frozen ECHO retrieval on LoCoMo categories 1–4 and LongMemEval-S, plus a matched 91-question QA sample against Mem0 OSS. They show that ECHO’s memory plane delivers high evidence coverage while its current reader underperforms Mem0 on strict QA accuracy.

BENCHMARK

Frozen ECHO Retrieval on LongMemEval-S and LoCoMo

Hit@10 on LoCoMo categories 1–4 and LongMemEval-S, plus matched QA accuracy for ECHO and Mem0 OSS.

BENCHMARK

BEAM-100K Retrieval Ablations for ECHO

Hit@10 and turn recall for full ECHO v11 versus dense-only, lexical-only, and no query expansion on the local BEAM-100K row.

KEY INSIGHT

The Counterintuitive Finding

On the BEAM development pilot, ECHO achieves 100% Hit@10 while recovering only 42.86% of annotated turns for event ordering and summarization.

This is counterintuitive because developers often assume high Hit@10 means sufficient evidence, but ECHO shows that one relevant hit can coexist with severely incomplete multi-event support.

WHY IT MATTERS

What this unlocks for the field

ECHO unlocks a governed memory plane where currentness, conflicts, and support coverage are explicit, auditable objects rather than hidden retrieval side effects.

Builders can now separate similarity-based candidate discovery from authoritative state resolution, design provenance-closed packers, and evaluate long-horizon agents at clear system boundaries instead of relying on opaque end-to-end scores.

~14 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

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

Answers use this explainer on Memory Papers.

Checking…