ECHO: An Auditable Dual-Path Memory System for Long-Horizon Agents

arXiv:2608.21755 · cs.AI · Submitted 2026-08-22 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "ECHO: An Auditable Dual-Path Memory System for Long-Horizon Agents".

Jane: Here's the gist of the paper. The gist: ECHO presents an auditable memory architecture and service prototype inspired by episodic encoding, consolidation, contextual reinstatement, reconsolidation, and executive control for long-horizon agents.

Tom: How it works ECHO follows a continual loop of encoding, organization, recall, and evolution immutable events retain episodes; projection builds typed revisions; hybrid routes propose candidates; a bitemporal ledger alone determines currentness; provenance closure restores the dependencies required by an operation.

Jane: This cognitive mapping motivates this decomposition but makes no anatomical, biological, or clinical claim.

Tom: The architecture separates discovery from authority by placing a bitemporal ledger and provenance closure between retrieval and realization.

Title and authors: Tom: So, what do these suggested improvements actually do? They seem like refinements to how you handle the core concepts of projection, provenance, and candidate discovery that we talked about earlier.

Jane: They focus heavily on making sure the ledger component doesn't just append revisions but actively manages them so that a revision can supersede or revoke a previous one without rewriting all the history. That’s a big deal for temporal integrity.

Lu: They are implementing this bitemporal ledger concept more rigorously, explicitly allowing revisions to be resolved in ways like superseding, revoking, or leaving another revision unresolved based on transaction time versus valid time.

Meng: I'm interested in how they formalize the candidate discovery process because that’s where I thought we could see a failure mode—the semantically close stale value outranking a current one. They are trying to solve that head-on.

Lalam: They introduce a worker-oriented retrieval pipeline where evidence packing uses specific sets, like an A count set for countable members and a temporal arithmetic set for both endpoints, ensuring internal completeness isn't mistaken for visible completeness.

Jane: That sounds like they are formalizing the context construction better so that what the system *thinks* it has in its context is directly related to what it *actually* needs to solve the current operation.

Tom: And then there’s this rule about semantic similarity not determining factual currentness; they state that the resolver must return either a state, or a miss, or a conflict, or a conflict set where the sadm function value is greater than one.

Lu: That's a very concrete engineering commitment because it directly addresses the failure mode we discussed earlier by making that separation explicit in the resolver’s output contract.

Meng: If they can guarantee that semantic proximity doesn't trick us into picking an outdated piece of information, that makes the whole retrieval stage much more trustworthy for production use.

Jane: And they tie this all together by defining surface contracts based on operation types—like requiring minimal sufficient support for a Fact operation—to ensure that the metrics we look at are truly diagnosing context construction rather than just guessing the answer.

Lalam: It sounds like they’re moving towards a system where the evidence and answer operations are tightly coupled with these budget and constraint checks, ensuring that every claim is backed by its own specific boundary.

Tom: So they're taking those cognitive inspirations—like reconsolidation or executive control—and translating them into specific architectural rules about how the data flows and how decisions are made at each stage.

Jane: It seems like this paper is less about proposing a completely new memory storage mechanism and more about proposing a verifiable, auditable framework on top of existing ideas that makes the process transparent.

Lu: The cognitive mapping serves to motivate this decomposition, but the engineering commitment is what drives it; it shows *why* we need these distinct components like projection and consolidation working separately.

Meng: For us, I see this as a blueprint for building memory layers where you can swap out the candidate discovery mechanism without having to rebuild the entire history management system from scratch.

Lalam: This focus on verifiable evidence boundaries seems to be what distinguishes this work, moving away from just a performance number toward a measurable contract about how the agent operates.

Jane: So, we’ve seen that ECHO is an auditable architecture and now we see it’s being improved by making the rules for revision and discovery much stricter and more verifiable.

The paper's summary: Tom: We’re coming to the wrap-up of this paper on ECHO: An Auditable Dual-Path Memory System for Long-Horizon Agents. So, what’s the main idea we need to remember about this architecture?

Jane: The main idea is that they've proposed a system that separates immutable experience from organized knowledge and uses a bitemporal ledger to manage revisions and currentness explicitly, making the entire process auditable through provenance closure.

Lu: It’s an architecture study whose claims are aligned with corresponding empirical boundaries, focusing on retrieval and context construction paths through stage-level evidence evaluation.

Meng: What this means for deployment is that we can build agents with much more reliable long-term memory because we have explicit controls over how information evolves and what is considered current.

Lalam: Their contribution is architectural and methodological: making the concepts of currentness, support, conflict, and failure observable while aligning each claim with its corresponding evidence boundary.

Tom: Yeah, they’re showing us exactly where the system can succeed or fail based on these measurable boundaries instead of just hoping for a high score.

Jane: It’s a functional design hypothesis about cognitive mapping that separates temporally situated experience from organized knowledge, and that's what makes this paper interesting to look at.

Lu: This approach shows how to use episodic encoding and reconsolidation principles in an engineering way, which is the big picture for long-horizon agents.

Meng: It’s a blueprint for building memory layers where you can swap out the candidate discovery mechanism without having to rebuild the entire history management system from scratch.

Lalam: The paper's contribution is architectural and methodological: making currentness, support, conflict, and failure observable while aligning each claim with its corresponding evidence boundary.

Tom: So we’re done with this deep dive into ECHO: An Auditable Dual-Path Memory System for Long-Horizon Agents. We’ll be taking a short break before we look at some of those other interesting papers from arXiv.

Jane: Right, stay tuned, because next up we’ve got a look at Domain Shift in Echocardiography.

The paper's improvements: Tom: So we just talked about how ECHO builds this architecture using episodic encoding and consolidation to handle long-term agent memory. Now we look at what they suggest to make that system actually work better in practice, which is where they introduce these specific improvements.

Jane: Right, so the paper isn't just describing a concept; it’s showing how you can take this framework and add concrete rules for its components, like how the ledger itself needs to function.

Tom: They are really focusing on making the bitemporal ledger more powerful. Instead of just appending data, they suggest that a revision can actually supersede or revoke an older one, which keeps the history clean without rewriting everything from scratch.

Jane: That’s important because it solves a problem where you might have conflicting information—you can’t just overwrite things blindly. It makes sure the current state is determined by the ledger, not just by how similar two pieces of text are to each other.

Tom: Exactly. Then they tackle candidate discovery, which is where we saw that failure mode earlier—where a semantically close old value might beat a current one. They’re proposing ways to formalize what evidence packing looks like using specific sets for counting and temporal arithmetic.

Jane: So it’s about making sure the system knows exactly what evidence it needs for a specific operation, not just grabbing whatever looks right based on similarity scores. It separates internal completeness from what you can actually see.

Tom: And they put a real constraint on the resolver itself. They say semantic similarity cannot decide factual currentness anymore; the resolver has to return a state, or a miss, or a conflict set where that sadness measure is greater than one.

Jane: That’s a big step for reliability because it creates an explicit barrier against that stale data problem we talked about. It forces the system to be honest about what it knows versus what it doesn't know.

Tom: Plus, they define surface contracts based on the operation itself, like for a Fact operation requiring minimal sufficient support and minimal supported answer. That connects the internal process directly to what’s actually needed for a task.

Jane: So they are moving from just building a memory structure to building an auditable service where every decision has a measurable boundary tied to its specific goal. It makes the system more trustworthy because you can check those invariants at every step.

Tom: That’s the core idea—making currentness, support, conflict, and failure observable through these strict rules. It shifts the focus from just getting a high retrieval score to proving that the evidence delivery is sound for the specific operation being asked.

Jane: And this architecture is really about providing a solid foundation so that when you build something on top of it—like an agent that needs to remember things across many long steps—you have something rigorously defined underneath it. So, what does all this mean for building those agents?

Conclusion: Tom: So we’ve gone through the details of ECHO: An Auditable Dual-Path Memory System for Long-Horizon Agents, and what I see is that this isn't just another memory trick; it’s a way to build verifiable trust into long-term AI agents by making the whole process auditable.

Jane: Right, so basically, they took those complex ideas from episodic encoding and consolidated them into a framework where you can actually see the evidence path, which is huge for building anything that needs to remember things over a long time.

Tom: Yeah, it’s about separating the experience—the raw data—from the knowledge structure by using that bitemporal ledger to track revisions without having to rewrite all the history every time something new comes in.

Jane: That distinction between temporal experience and organized knowledge is what really sets this apart for how we think about agents that have to operate over many steps.

Lu: I think the way they formalized the candidate discovery process, using those specific sets for packing evidence, really shows how they were reaching for a concrete engineering solution to a problem where similarity scores just weren't enough.

Meng: From an engineering standpoint, I’m interested in that constraint on the resolver—that it can’t just pick something because it looks semantically close; that rule about the sadness measure being greater than one is what keeps me up at night when we think about deploying this at scale.

Lalam: From my perspective, I see this as a much more reliable way for an AI to ground its answers because it’s not just guessing based on the most similar text; it’s checking against the current validity defined by that ledger. I’m very sure of its answer here because the framework itself is so rigorous.

Tom: The implication is that instead of deploying a black box memory system, we get something where we can actually audit why an agent made a certain decision or retrieved a piece of information.

Jane: It means that for any application—say, an AI assistant or a complex autonomous system—you have measurable boundaries for when its memory is current and what evidence it's relying on.

Lu: The future work they hint at involves making sure these architectural components are modular so you can swap out the discovery mechanism without breaking the whole system, which makes sense for iterative research.

Meng: But I’d ask, how much overhead does all this provenance closure and ledger checking add to the runtime? It has to be efficient enough to actually run on a server instead of just being a neat theoretical model.

Lalam: I think the cost of that overhead is worth it if it guarantees we aren't getting misled by stale information, which is what I’m focused on here.

Tom: So for anyone listening, remember ECHO: An Auditable Dual-Path Memory System for Long-Horizon Agents is about building a memory system where you can prove the evidence path and control the revision process.

Jane: It’s a solid piece of architecture that gives us much better tools to build agents that actually handle long horizons without just hallucinating or relying on outdated facts.

Tom: Next up, we’re looking at Domain Shift in Echocardiography, where we see how domain change affects medical image processing.

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

XDream Robotics

cs.AI

Submitted: 2026-08-22

Updated: 2026-10-20

Comments: 8 pages. Revised title and dual-path architecture; updated five-system aggregate evaluation and clarified reader configuration, evidence scope, and limitations

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: The gist: ECHO presents an auditable memory architecture and service prototype inspired by episodic encoding, consolidation, contextual reinstatement, reconsolidation, and executive control for

Key concepts

Bitemporal Ledger: This is a ledger that tracks both valid time (when an event actually happened) and knowntime (when the system currently knows about it). A key feature is that a revision can supersede, revoke, or leave another revision unresolved. This ensures the current state is determined by the ledger's rules, not just similarity scores.
Provenance Closure: This mechanism restores all necessary dependencies required to perform an operation. It ensures that when retrieving evidence for a result, you get not just the answer but also all the links and preceding events needed to verify that result. This is crucial for auditability and correctness.
Candidate Discovery: This process combines several methods—typed routes, lexical matching, semantic retrieval, and neighborhood expansion—to propose potential pieces of evidence. The system then uses the bitemporal ledger to filter these candidates based on their currentness and revision status. This prevents stale or conflicting information from being incorrectly selected as the truth.
Evidence Closure (Admissible Context): This is a formal method for selecting a budget of evidence units that are sufficient to verify an answer. It involves packing evidence such that the total size stays within a budget while satisfying specific constraints related to the operation. This ensures internal completeness aligns with what is actually needed for the task.

Summary

The gist: ECHO presents an auditable memory architecture and service prototype inspired by episodic encoding, consolidation, contextual reinstatement, reconsolidation, and executive control for long-horizon agents.

How it works

ECHO follows a continual loop of encoding, organization, recall, and evolution immutable events retain episodes; projection builds typed revisions; hybrid routes propose candidates; a bitemporal ledger alone determines currentness; provenance closure restores the dependencies required by an operation. This cognitive mapping motivates this decomposition but makes no anatomical, biological, or clinical claim.

The architecture separates discovery from authority by placing a bitemporal ledger and provenance closure between retrieval and realization. Lexical and dense similarity are effective ways to propose what should be inspected, but they are poor mechanisms for deciding which revision is current or whether two states conflict.

Cognitive Inspiration and Engineering Commitments

The cognitive mapping motivates this decomposition but makes no anatomical, biological, or clinical claim. The episodic–semantic distinction separates temporally situated experience from organized knowledge. Complementary learning systems explain why rapid experience capture and slower structured learning serve different roles. Reconsolidation shows that recall can reopen an established memory to revision. Temporal-context models connect recall to a changing internal context. Cognitive-control accounts emphasize goal-dependent selection of what guides behavior.

Target Runtime and Evaluation Scope

The target runtime and evaluation scope are defined by decoupling durable ingestion from optional projection and index maintenance. The empirical study evaluates durable carrier, candidate discovery, and context-construction paths at service level, and assesses V25 ledger, provenance closure, operation-planning, and surface-contract properties through domain-level tests.

Projection, Provenance, and Revision Integrity

Projection converts raw turns and events into typed units while preserving raw source pointers. The ledger component appends revisions rather than overwriting prior state. A revision may supersede, revoke, or leave another revision unresolved, allowing late-arriving evidence to change transaction-time knowledge without rewriting historical valid time. The resolver behavior is evaluated through component-level invariants.

Candidate Discovery and Currentness Authority

The candidate union combines typed routes, lexical matching, semantic retrieval, and fixed neighborhood expansion. Fusion produces a candidate set, not truth. The ledger subsequently applies the requested valid/knowntime view and revision status. This separation prevents a frequent failure: a semantically close stale value outranking a less similar current value.

Evidence and Answer Operations

A retrieval hit is not necessarily sufficient evidence. The evidence closure Γ(u) contains the originating events, required revision links, and dependencies needed to verify the result. An admissible context is then an atomic-closure packing problem s.t. X u∈H xu tokens(Γ(u)) ≤ B, for utility-conditioned evidence units H and budget B. The public question induces an operation such as FACT, COUNT, LIST, TEMPORALLOOKUP, or TEMPORALARITHMETIC.

Objective and Claim Taxonomy

The reader computes z = D(q, C, o), a = R(q, z, h). The system objective is therefore constrained rather than a single retrieval score. For evaluation rows i = 1..., n, we report J = 1/n Xn i=1 [ai ≡ yi] − λsLstale − λcLconflict − λfLfalse abstain − λbB, subject to provenance closure and generation-fencing invariants.

Failure Semantics

Every public row and every arm-budget cell is retained. Reader transport, parsing, and judge failures score zero. Exact abstentions are scored deterministically; remaining bytedistinct answers are judged once per question-response pair and reused across consuming cells.

Limitations and Ethical Considerations

The current study evaluates evidence delivery, not autonomous behavior or human-like cognition. The cognitive mapping is a functional design hypothesis, not a claim that software modules are anatomically, neurally, or clinically equivalent to biological memory. Persistent personal memory raises privacy, consent, deletion, and access-control risks. The paper claims an auditable architecture, component-level invariants, and development-stage protocol findings.

Conclusion

ECHO is an auditable memory architecture and service prototype separating immutable experience, bitemporal authority, candidate discovery, provenance closure, and realization. Its visible-dataset retrieval scores are descriptive because sourcespecific expansion rules were used; it also fails the fresh BEAM gate, and Mem0 OSS is more accurate on the fixed 91-question QA sample. The contribution is therefore architectural and methodological: make currentness, support, conflict, and failure observable, while aligning each claim with its corresponding evidence boundary.

Appendix A

All reported ECHO retrieval rows must be bound to dataset, source, model, prompt, and executable hashes. Retrieval and QA metrics are separate estimands.

Appendix B

Frozen ECHO artifacts include the LoCoMo v7, LongMemEval v5, BEAM v11 development pilot, and five-history fresh BEAM gate summaries and question-level audits. The evidence ledger also includes four complete BEAM ablation pairs, their analyses, and the 40-configuration worker-stress report bundle. Frozen ECHO artifacts include the LoCoMo v7, LongMemEval v5, BEAM v11 development pilot, and five-history fresh BEAM gate summaries and question-level audits. The artifact ledger is organized around the associated paths, datasets, prompts, images, container revisions, Qwen model IDs, environment metadata, and failures by SHA-256.

References

[1] A. Maharana et al., “Evaluating very long-term conversational memory of LLM agents,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024, pp. 13 851–13 870

[4] K. Torp et al., “Effective timestamping in databases,” The VLDB Journal, vol. 8, no. 3–4, pp. 267–288, 2000

[14] P. Chhikara et al., “Mem0: Building production-ready AI agents with scalable long-term memory,” in European Conference on Artificial Intelligence, 2025, pp. 2993–3000

[16] C. Packer et al., “MemGPT: Towards LLMs as operating systems,” arXiv preprint arXiv:2310.08560, 2023

[18] P. Rasmussen et al., “Zep: A temporal knowledge graph architecture for agent memory,” arXiv preprint arXiv:2501.

Improvements for AI systems

Evaluation Protocol Improvement: Adopt a strict rule that The reader may enumerate members, compare revisions, or normalize temporal endpoints internally, but the public result is R(D(q, C, Io), Fo). This ensures that internal processing steps are separated from the final output contract to prevent conflating retrieval quality with answer correctness.
Memory Architecture Upgrade: Implement a bitemporal ledger where a revision may supersede, revoke, or leave another revision unresolved. This allows the system to explicitly handle revisions without rewriting history, ensuring that the currentness owner is determined solely by the ledger, not by similarity scores.
Retrieval Pipeline Refinement: Integrate a worker-oriented retrieval pipeline where evidence packing uses sets defined as A count set groups the countable members; a temporal arithmetic set groups both endpoints; a current-state set groups the active revision with the revision evidence needed to exclude an older value. This formalizes how context is constructed based on operation requirements, ensuring internal completeness is not equivalent to visible completeness.
Decision Boundary Enforcement: Ensure that Semantic similarity cannot determine factual currentness, by explicitly enforcing that the resolver returns either a state, ⊥miss, ⊥conflict, or a conflict set where "Sadm(k, t) > 1. This prevents the failure mode where a semantically close stale value outranking the current revision."
Operational Constraint Alignment: Define surface contracts based on operation type using Operation Internal requirement Final surface, such as for Fact operations requiring Minimal sufficient support Minimal supported answer, ensuring that retrieval metrics diagnose context (C) and do not substitute for correctness (a).

Abstract

Long-horizon agents require more than retrieval over stored conversations: they must preserve source evidence, revise outdated knowledge, and distinguish relevant memories from sufficient support for an answer. We present ECHO (Embodied Context & History Orchestration), an auditable dual-path memory system. A fast path captures immutable events and serves multi-route recall, while asynchronous workers consolidate experience into revisable state and rebuild derived indexes. Source provenance, temporal authority, and answer realization are explicit boundaries. The service supports original-source packets with evidence-use checks and disables model-based reranking by default. For the latest reported accuracy, the authors confirm mixed retrieved inputs comprising original dialogue and memory summaries, answered by Qwen3-8B with NF4 quantization. In the latest author-supplied five-system aggregate evaluation, ECHO reports 56.27% answer accuracy, 98.04% source Hit@10, 93.58% macro source Recall@5, and 13.92% hallucination flags, with 19.54 ms mean retrieval latency. Its reported accuracy exceeds Mem0 by 4.53 percentage points in this snapshot. The results are descriptive aggregates; matched question-level evaluation and generation latency remain to be established. We separate these observations from earlier published experiments and engineering regressions.