# Source ledger

Latest source update: **2026-09-20**. AriGraph's update mechanism, Zep's temporal model and outcomes, LongMemEval's retrieval/QA criteria, SLoD's evaluation scope and the graph-memory survey's scope were checked against the cited full paper versions on this date. MemOS and Harness the Memory retain their September 19 check dates.

This ledger records where each field group came from. Every row cites the full paper sections used to interpret its evaluation scope. Hugging Face pages remain the discovery links; `metadata_source` pins arXiv bibliographic metadata where the HF author list lags the reviewed revision. GitHub repository metadata provides the code-repository license identifier. A repository license is not assumed to cover the paper, dataset, model, or every dependency.

The MIT license covers original files in this Space. Linked papers, repositories, datasets, and project pages remain under their own licenses and terms; links and factual summaries do not relicense upstream works.

## Evidence rows

| Artifact | Scope, claims, and limitations | Code / project | License evidence | Fields intentionally left unknown |
|---|---|---|---|---|
| [Semantic Level of Detail for Knowledge Graphs: Discovering Abstraction Boundaries via Spectral Heat Diffusion](https://huggingface.co/papers/2603.08965) | [arXiv v2 sections 3, 6 and 8](https://arxiv.org/html/2603.08965v2) · abstraction-boundary experiments on HSBM and WordNet; query-to-scale selection and end-to-end QA are future work | [Public code companion](https://github.com/mnemoverse/mnemoverse-slod-paper), identified independently of the paper API's absent repository field | GitHub API reports Apache-2.0; pinned [LICENSE](https://github.com/mnemoverse/mnemoverse-slod-paper/blob/062a68876d4eb3ffdeaedb1459f2eb4e9955434e/LICENSE) and [NOTICE](https://github.com/mnemoverse/mnemoverse-slod-paper/blob/062a68876d4eb3ffdeaedb1459f2eb4e9955434e/NOTICE) checked for the code repository only | time model, conflict/update model, retrieval/answer/agent-task outcome class, dataset URL; public code availability does not establish independent replication |
| [LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory](https://huggingface.co/papers/2410.10813) | [arXiv v2 sections 3.3, 5 and E.5](https://arxiv.org/html/2410.10813v2) · Recall@k/NDCG@k and LLM-judged QA; some correct knowledge-update answers fail the strict retrieval criterion when earlier evidence is missing | [Official repository](https://github.com/xiaowu0162/LongMemEval) links the [cleaned dataset](https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned); its GitHub data directory contains a sample generator, not the benchmark instances | GitHub API reports MIT for code; the [pinned cleaned-dataset card](https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/blob/98d7416c24c778c2fee6e6f3006e7a073259d48f/README.md) also declares MIT | broad agent-task transfer; the cleaned replacement is not a byte-identical original evaluation snapshot |
| [Zep / Graphiti](https://huggingface.co/papers/2501.13956) | [arXiv v1 sections 2.1, 2.2.3 and 4](https://arxiv.org/html/2501.13956v1) · bi-temporal edges, LLM-assisted invalidation with ingestion-order priority; DMR and LongMemEval measure answer quality, not retrieval-only quality | [Graphiti repository](https://github.com/getzep/graphiti) | GitHub API reports Apache-2.0 for the code repository | exact experiment dataset URL; independent replication; lossless history reconstruction; stale-source reimport |
| [AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents](https://huggingface.co/papers/2407.04363) | [arXiv v3 sections 2, 4 and 5](https://arxiv.org/html/2407.04363v3) · new triplets are compared with existing semantic edges and outdated edges are removed; TextWorld and NetHack plus MuSiQue/HotpotQA give mixed outcomes. [v3 metadata](https://arxiv.org/abs/2407.04363v3) adds Andrey Kravchenko | [AriGraph repository](https://github.com/AIRI-Institute/AriGraph) | GitHub API reports MIT for the code repository | dataset URL; preservation and historical retrieval of conflicting versions; independent replication |
| [MemOS](https://huggingface.co/papers/2507.03724) | [arXiv v4 §6](https://arxiv.org/html/2507.03724v4) · LoCoMo, LongMemEval, PreFEval, PersonaMem answer evaluations; retrieval/API performance and KV studies are separate. [v4 metadata](https://arxiv.org/abs/2507.03724v4) governs authors | [MemOS repository](https://github.com/MemTensor/MemOS) · [project page](https://memos.openmem.net/) | GitHub API reports Apache-2.0 for the code repository | dataset URL; independent retrieval-quality outcome; agent-task completion; conflict preservation; independent replication |
| [Graph-based Agent Memory](https://huggingface.co/papers/2602.05665) | [arXiv v1 sections V to VIII and XI](https://arxiv.org/html/2602.05665v1) · taxonomy, evolution and existing benchmark survey; topical coverage is not a new experiment | [Awesome GraphMemory](https://github.com/DEEP-PolyU/Awesome-GraphMemory) · curated resource repository, not an implementation repository | GitHub API returned no SPDX license identifier | implementation code, dataset URL, evaluated outcome level, common time/conflict model |
| [Harness the Memory](https://huggingface.co/papers/2608.15008) | [arXiv v1 sections 3 to 5, appendices C and G](https://arxiv.org/html/2608.15008v1) · task regimes, retrieval depth and history length; QA and agent tasks; most-recent-fact selection is not contradiction preservation | Code/harness release unverified; the paper promises code upon acceptance | unknown; no upstream repository license inferred | code, harness dataset, released adaptive router, independent replication |

## Interpretation rules

- “Available” means a verified public URL exists; it does not certify completeness or reproducibility.
- “Unknown” means the checked primary source did not verify the field.
- “Not evaluated” must be checked against the full paper's evaluation/experiments, not inferred from an abstract's omission. Surveys can provide context without an evaluation result of their own.
- Reported results remain bound to each paper’s original task and setup.
- `Answer` includes QA and personalization quality. `Mixed` requires more than one evaluated outcome class (retrieval, answer, or agent task); latency, cost and API success alone do not create another outcome class. Zep v1 and MemOS v4 therefore have `answer`, while AriGraph v3 combines QA with interactive outcomes and has `mixed`.
- Reported evaluations are author-reported; independent replication is not established by this source review. This limitation applies equally to Mnemoverse-authored work and other systems.
- Evidence labels describe the cited paper scope, not all results or capabilities in the project's current codebase. Later repository results do not silently become results of the mapped paper.
- SLoD is labeled Mnemoverse-authored because Edward Izgorodin is the sole listed author.

## Platform references

- [Spaces overview](https://huggingface.co/docs/hub/en/spaces-overview)
- [Static HTML Spaces](https://huggingface.co/docs/hub/spaces-sdks-static)
- [Spaces configuration](https://huggingface.co/docs/hub/main/spaces-config-reference)