Pith.

Artificial Intelligence · arXiv:2611.21998 v1 · submitted 16 Nov 2026 · 57 references

KV-LTM: Persistent Per-Session KV Caches as Long-Term Memory for Personal Assistants

Ziv Kfir, Ido Kfir

The paper says

“Personal assistants accumulate months of conversations, but today's long-term memory stacks put derivation or selection between a question and its evidence.”

more of the abstract

In a system-level comparison on LongMemEval, two commercial assistants scored well below the same model reading the history offline [1], consistent with, but not isolating, content loss between write and read. We propose KV-LTM: a memory tier storing each session's raw text plus its KV cache under a fixed reader model, answering by restoring cached sessions and appending the question. KV-LTM is complementary to, not a replacement for, retrieval and indexing. Text is the memory of record; the cache is a recompute-free prefill accelerator, degrading gracefully to text. We specify the design: storage record, budget-derived retention, append-only read path, and bounded aggregation. We size it analytically: across pinned BF16 full-attention (GQA) and latent-cached MLA configurations, per-token state spans approximately 31K-328K bytes (the pinned KDA+MLA hybrid differs: approximately 8.1K bytes/token plus a fixed per-session recurrent term), giving approximately 0.34-3.6 TB per user-year at a declared 30K-token/day scenario, against approximately 49 MB of raw text; per-query cost is modelled as regime-conditional transfer, attention, decode, aggregation, and compute. We define an evaluation protocol needing no persistent-cache implementation for logical QA evaluation: per-session prefill emulates valid cache reuse under the same model/input-state contract and ideal arithmetic, with named arms, controlled baselines, and a preregistered reader-model gate; a separate minimal real-cache replay tests numerical equivalence and measures the cost axis, with outcomes not assumed. KV-LTM trades storage for avoided re-prefill compute and write-time fidelity, within a declared retention budget.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 18 Nov 2026 · read the full text (118,042 characters) · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. KV-LTM is a detailed design proposal for a persistent per-session KV-cache tier for personal assistants, with analytic cost models and a thorough evaluation protocol, but it contains no experimental results and its central feasibility claims rest on unvalidated assumptions about user behavior and reader sufficiency. This line is wrong

9 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

The paper, first character to last

The reviewer was given the whole text, 118,042 characters. Not checked: figures, code, what the 57 references say.

9 minor objections

The paper says, at §1.1

Pith says: minor objection 1 of 9

The L1-L4 taxonomy is introduced but not used consistently throughout the paper. It appears in Figure 1 and a few other places, but the paper often refers to 'reference path', 'hybrid variant', etc. without mapping back to L4. Consider either using the taxonomy more systematically or removing it to avoid confusion.

The paper says, at §2.2

Pith says: minor objection 2 of 9

The exactness lemma is stated clearly, but the paper later notes that practical equality must be verified with tolerances by a real-cache replay (§8). This is appropriate, but the lemma's conditions are quite restrictive (same model, same prefix, same positions, etc.). The paper should discuss how often these conditions hold in practice, e.g., when the model is updated or the prompt template changes.

show 7 more

The paper says, at §4.6

Pith says: minor objection 3 of 9

The aggregation contract specifies bounds S, e, o but does not give example values or a rationale for how they are chosen. The paper would benefit from a brief discussion of how these parameters affect answer quality and cost.

The paper says, at §5.2

Pith says: minor objection 4 of 9

The table in §5.2 lists MiniCPM3-4B as having 62 layers. Checking the cited HuggingFace config (openbmb/MiniCPM3-4B) shows 62 layers, but this is unusually deep for a 4B model. The paper should verify this number or note that it comes from the config.

The paper says, at §6.2

Pith says: minor objection 5 of 9

The text-only contrast is useful, but the paper should clarify whether this is the raw text of the sessions only, or includes metadata and other overhead. The storage record in §4.1 includes text, state, meta, serialization, and digests, so the total storage is larger than the text-only figure.

The paper says, at §7.2

Pith says: minor objection 6 of 9

The worked examples in §7.2 assume 2,000-token sessions and 64 output tokens per session. These are plausible but arbitrary. The paper should discuss how results change with different session lengths and output token counts, or provide a sensitivity analysis.

The paper says, at §8.5

Pith says: minor objection 7 of 9

The paper states that the <70B-class reader suffices hypothesis is gated, but the gate is not fully specified. What are the 'task-quality and abstention targets relative to a named larger-reader baseline'? What is the numeric O-dev accuracy tolerance Δ? These should be preregistered before the evaluation runs.

The paper says, at §9.2

Pith says: minor objection 8 of 9

The privacy discussion is brief and acknowledges limitations. However, the paper does not discuss the implications of storing KV caches, which may encode information about the session content in a way that is not easily redactable. The paper notes that no state-to-text inversion is claimed, but this is an important area for future work.

The paper says, at Appendix A

Pith says: minor objection 9 of 9

The appendix is thorough and clearly labels assumptions. However, some derivations are missing intermediate steps. For example, the Kimi-Linear hybrid formula 'sessions × S_recurrent + tokens × b_MLA' is stated but S_recurrent is left symbolic, making it impossible to compute total storage for this configuration.

What would settle it

If a controlled experiment using the paper's evaluation protocol (arm A: per-session prefill + aggregation) on a benchmark like LongMemEval or LoCoMo shows that the <70B reader model achieves accuracy no better than or worse than existing memory systems (e.g., Mem0, Zep) on questions requiring evidence from multiple sessions, then the central claim that KV-LTM provides write-time fidelity and avoids content loss would be undermined, as the reader's limited capability would negate the benefit of preserving raw content.

from Pith's summary of the paper

How the review went

5 passes by deepseek-v4-flash · 18 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic concern withdrawn: the paper already says this 7 s · 39,298 tokens read, 637 written
  3. Referee 9 minor 18 s · 39,541 tokens read, 2,464 written
  4. Check against the paper kept 0 of 5 objections 27 s · 80,691 tokens read, 5,200 written
  5. Editor wrote the verdict above 7 s · 39,936 tokens read, 572 written

Not run on this paper: the simulated authors' defence.

Pith's passes raised 5 objections. A check against the paper's text kept 0 and withdrew 5 (5 the paper already says).

The review in full

The review in brief, machine-written

This paper proposes KV-LTM, a memory tier that stores each session's raw text plus its KV cache under a fixed reader model, answering questions by restoring cached sessions and appending the question. The core idea is that the KV cache acts as a recompute-free prefill accelerator, with text as the memory of record that degrades gracefully when the cache is evicted or invalidated.

The paper does several things well. It clearly distinguishes KV-LTM from prior work like CAG and prefix caching, positioning its contribution as the per-session retention/eviction policy, read/aggregation protocol, PA-scale cost model, and evaluation protocol rather than a new memory primitive. The design specification is thorough, covering storage records, cache validity contracts, write and read paths, aggregation, retention, invalidation, and fallback. The analytic cost models are detailed and transparent about their assumptions, providing per-token KV sizes across multiple model configurations and annual storage envelopes under declared volume scenarios. The evaluation protocol is well-structured, with named arms, controlled baselines, and a preregistered reader-model gate, and it cleverly uses per-session prefill as a stand-in for cache reuse to enable evaluation without implementing persistent caching.

The soft spots are significant. The paper contains no experimental results or implementation. The central feasibility claims rest on declared assumptions (30K tokens/day, 2,000-token sessions, 5 GB/s bandwidth) that are not validated against real PA usage data. The reader-sufficiency hypothesis (that a <70B-class reader suffices for the retained-window read path) is explicitly acknowledged as unproven. The cost models are analytic estimates, not measured latencies. The evaluation protocol is well-specified but has not been executed. The paper is transparent about these limitations, which is commendable, but the lack of empirical validation means the design's practicality remains an open question.

This paper is for researchers working on LLM memory systems for personal assistants. It provides a clear design and a testable evaluation protocol. The paper deserves a serious referee, but the referee should focus on whether the assumptions are reasonable and whether the evaluation protocol is sufficient to validate the claims. The paper would be strengthened by measuring real PA usage to validate volume assumptions, implementing the evaluation protocol and reporting results, testing the reader-sufficiency hypothesis, and providing a minimal real-cache replay to verify the exactness lemma in practice.

The stress test's concern

Withdrawn: the paper already says this.

The paper says

“\item \textbf{The reader-sufficiency hypothesis is unproven at drafting.} §8.5 states and gates the ``\textless70B-class reader suffices'' hypothesis; it remains a hypothesis until the preregistered gate runs (§8.5, §8.7).”

The referee's report

Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 4 of its 4 major objections did not survive that check.

Pith's reading of the paper (a machine's summary, not the authors')

The paper argues that current long-term memory systems for personal assistants lose content either at write time (by deriving summaries or facts) or at read time (by selecting a subset via retrieval). The authors propose KV-LTM, a memory tier that stores each session's raw text plus its native KV cache under a fixed reader model. At query time, the system restores cached sessions and appends the question, avoiding repeated prefill of session text while preserving the original content. The paper provides analytic cost models showing per-token KV state spans approximately 31K–328K bytes, giving approximately 0.34–3.6 TB per user-year at a declared 30K-token/day scenario, against approximately 49 MB of raw text. The design is complementary to retrieval and indexing, and the paper defines an evaluation protocol that uses per-session prefill as a stand-in for cache reuse, avoiding the need for a full persistent-cache implementation for logical QA evaluation.

What carries the argument. The key machinery is the per-session KV cache as a recompute-free prefill accelerator. Under the exactness lemma (Section 2.2), restoring a cached session and appending a question is mathematically identical to a full prefill of (session ‖ question) under ideal arithmetic, provided the same model, token prefix, position indices, RoPE configuration, attention masks, and native cache state are preserved. This allows the system to avoid re-prefilling session text at query time, while the raw text remains the memory of record for fallback.

Core claim. The paper claims that storing per-session KV caches alongside raw text, and restoring them at query time, provides a recompute-free prefill accelerator that avoids write-time derivation loss and query-time selection loss, within a declared retention budget. The central claim is that this design is feasible for personal assistants, with storage costs of 0.34–3.6 TB per user-year depending on model architecture, and that the trade-off of increased storage for avoided re-prefill compute and write-time fidelity is worthwhile. The paper does not present experimental results but provides an evaluation protocol to test the logical read path without implementing persistent caching.

The premise it rests on. The paper assumes that a <70B-class reader model suffices for the retained-window read path, i.e., that per-session reading from cached KV states with a small model can reliably answer questions from session evidence, which is stated as a hypothesis to be tested, not a proven claim.

If the paper is right (Pith's inferences, not the authors')

If KV-LTM works as claimed, personal assistants could answer questions about past conversations without the content loss observed in current memory systems (e.g., Mem0, Zep, rolling summaries), potentially improving accuracy on long-term memory benchmarks.

The design enables a bounded recent evidence-preserving tier that reads retained sessions directly without a retrieval index, and a verifier/reranker role behind existing retrieval systems.

The analytic cost models suggest that for a 512 GB storage budget, achievable retention ranges from ~4 months (Llama-3.1-8B) to ~18 months (DeepSeek-V2-Lite), making the design practical for consumer hardware.

The evaluation protocol allows measuring the logical read path on public benchmarks without implementing persistent caching, using per-session prefill as an exact stand-in, which could accelerate research into KV-cache-based memory systems.

The paper's emphasis on write-time fidelity suggests that any system that derives or compresses session content at write time (summaries, extracted facts, graphs) will lose information that could be relevant to future questions. This implies that verbatim retention, combined with efficient read-time access, may be a more robust approach for personal assistants.

The cost models show that KV-LTM's storage requirements are dominated by the model architecture (GQA vs MLA vs hybrid), not by the volume of text. This suggests that future model designs with smaller per-token KV state (e.g., MLA, hybrid architectures) could make persistent KV caching more practical.

The paper's evaluation protocol, which uses per-session prefill as a stand-in for cache reuse, could be adopted by other researchers to evaluate similar memory designs without building full infrastructure, potentially accelerating progress in this area.

Check this review

What was read

The passes were given all 118,042 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): e4c205497459fca3c435b6057a160cac3833fb43eae7f463ea842751ffa77888

Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:

PYTHONPATH=. python scripts/review_input_check.py 2611.21998

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.014 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.0037 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.0040 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Check against the paper deepseek-v4-flash · medium effort · $0.011 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
  5. Editor deepseek-v4-flash · medium effort · $0.0037 · instructions of the Editor, fingerprint e7211f6a3d27…
  6. Circularity check deepseek-v4-flash · medium effort · $0.0037 · instructions of the Circularity check, fingerprint e9946e39d67b…
  7. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0040 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.044 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00

Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.

curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum

prints the fingerprint of the reader's instructions. The same works for every pass.

The receipt

The review's fingerprint was taken on 18 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): 24268a8722019d4520f7f8b5d766e7be9622c4f246b2abc084a55f64b5283ec0

curl -s 'https://pith.science/api/v1/verdicts/2611.21998?form=canonical' | sha256sum

prints the fingerprint of the review as it stands now. The rule, in four steps.

A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Personal assistants accumulate months of conversations, but today's long-term memory stacks put derivation or selection between a question and its evidence. In a system-level comparison on LongMemEval, two commercial assistants scored well below the same model reading the history offline [1], consistent with, but not isolating, content loss between write and read. We propose KV-LTM: a memory tier storing each session's raw text plus its KV cache under a fixed reader model, answering by restoring cached sessions and appending the question. KV-LTM is complementary to, not a replacement for, retrieval and indexing. Text is the memory of record; the cache is a recompute-free prefill accelerator, degrading gracefully to text. We specify the design: storage record, budget-derived retention, append-only read path, and bounded aggregation. We size it analytically: across pinned BF16 full-attention (GQA) and latent-cached MLA configurations, per-token state spans approximately 31K-328K bytes (the pinned KDA+MLA hybrid differs: approximately 8.1K bytes/token plus a fixed per-session recurrent term), giving approximately 0.34-3.6 TB per user-year at a declared 30K-token/day scenario, against approximately 49 MB of raw text; per-query cost is modelled as regime-conditional transfer, attention, decode, aggregation, and compute. We define an evaluation protocol needing no persistent-cache implementation for logical QA evaluation: per-session prefill emulates valid cache reuse under the same model/input-state contract and ideal arithmetic, with named arms, controlled baselines, and a preregistered reader-model gate; a separate minimal real-cache replay tests numerical equivalence and measures the cost axis, with outcomes not assumed. KV-LTM trades storage for avoided re-prefill compute and write-time fidelity, within a declared retention budget.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-18

The authors' main claim, as the classifier picked it out 0.86

“KV-LTM trades storage for avoided re-prefill compute and write-time fidelity, within a declared retention budget.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.91 · human participants 0.05 · controlled comparison 0.95 · randomized assignment 0.03 · states quantitative result 0.97 · compares with prior results 0.95 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Personal assistants accumulate months of conversations, but today's long-term memory stacks put derivation or selection between a question and its evidence. background 0.71 0.06 0.04 asserted 1.00
In a system-level comparison on LongMemEval, two commercial assistants scored well below the same model reading the history offline [1], consistent with, but not isolating, content loss between write and read. background 0.44 0.29 0.05 hedged possibility 0.72
We propose KV-LTM: a memory tier storing each session's raw text plus its KV cache under a fixed reader model, answering by restoring cached sessions and appending the question. contribution 0.76 0.54 0.19 asserted 0.99
KV-LTM is complementary to, not a replacement for, retrieval and indexing. limitation 0.65 0.24 0.08 asserted 1.00
Text is the memory of record; the cache is a recompute-free prefill accelerator, degrading gracefully to text. interpretation 0.51 0.37 0.18 asserted 1.00
We specify the design: storage record, budget-derived retention, append-only read path, and bounded aggregation. method 0.84 0.83 0.06 asserted 0.98
We size it analytically: across pinned BF16 full-attention (GQA) and latent-cached MLA configurations, per-token state spans approximately 31K-328K bytes (the pinned KDA+MLA hybrid differs: approximately 8.1K bytes/token plus a fixed per-session recurrent term), giving approximately 0.34-3.6 TB per user-year at a declared 30K-token/day scenario, against approximately 49 MB of raw text; per-query cost is modelled as regime-conditional transfer, attention, decode, aggregation, and compute. method 0.41 0.90 0.06 asserted 0.97
We define an evaluation protocol needing no persistent-cache implementation for logical QA evaluation: per-session prefill emulates valid cache reuse under the same model/input-state contract and ideal arithmetic, with named arms, controlled baselines, and a preregistered reader-model gate; a separate minimal real-cache replay tests numerical equivalence and measures the cost axis, with outcomes not assumed. method 0.97 0.82 0.06 asserted 0.98
KV-LTM trades storage for avoided re-prefill compute and write-time fidelity, within a declared retention budget. interpretation 0.88 0.57 0.55 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.36

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-18: 1 request, 7,156 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (4)
Figure 1 of the paper
Figure 1. Published system-level comparisons under two different benchmarks and source evaluation protocols: (a) LongMemEval’s 97-question subset with 3–6-session histories [1, Fig. 4a]; (b) LoCoMo J scores from Mem0’s vendor-authored evaluation [2]. Full-history reading scores highest among the selected systems shown in each panel, not across all evidence. Zep’s separate vendor-reported LongMemEval-S com￾parison favours its retrieval system over a full-context baseline averaging about 115K tokens and is not plotted here (§3.2, §8.1). These comparisons do not isolate a mem￾ory mechanism, and…
Figure 2 of the paper
Figure 2. KV-LTM write flow (once per session): canonical per-session prefill creates the stored state at session end; session text, metadata with integrity digests, and native reader state are retained under the §4.8 storage and latency policy, which may evict or invalidate state — the text is always kept. Boxes denote logical steps or stored objects, not software components. 4. On failed source verification: recover a known-good copy and re-verify the re￾covered source; on success, fall through to the state-validity check against the recovered text/serialization (load compatible state or…
Figure 3 of the paper
Figure 3. KV-LTM read flow (per question): each retained session’s text is verified against its stored digest first — a known-good copy is recovered and re-verified on failure, and an unrecoverable session counts as unavailable (out of coverage), never as evidence. Valid state is read from a request-local working copy; when policy al￾lows, invalid or evicted state is rebuilt from the verified text (§4.3), otherwise the session is counted as not cached and delegated to the §4.8 text fallback, outside the depicted cache-read path. Per-session outputs supply bounded evidence to the aggre￾gator…
Figure 4 of the paper
Figure 4. BF16 state bytes per token, logarithmic scale, for six pinned configura￾tions and one representative DeepSeek-V3-class assumption (marked †, not a pinned checkpoint). MLA values assume compressed-latent storage. Kimi-Linear shows only its MLA-layer component (bMLA = 8,064 B/token); its per-session recurrent state is symbolic and not drawn. The plotted per-token spread is ∼40× (the hybrid’s bMLA part at the low end); among the non-hybrid configs the spread is ≈10.5× (31,104 → 327,680). 6.1 Volume scenarios One coverage parameter drives everything: the fraction of a user’s daily…

Figures are the authors', reproduced from the paper.

References (57)
  1. [1] LongMemEval (Wu et al., ICLR 2025). https://arxiv.org/abs/2410.10813 https://github.com/xiaowu0162/longmemeval 42 on PithPith's reviewarXiv
  2. [2] Mem0. https://arxiv.org/abs/2504.19413 Vendor-authored paper (ECAI 2025) on PithPith's reviewarXiv
  3. [3] Zep. https://arxiv.org/abs/2501.13956 Vendor-authored preprint on PithPith's reviewarXiv
  4. [4] MemPalace. https://mempalace.net Project + independent analysis https:// arxiv.org/abs/2604.21284 on PithPith's reviewarXiv
  5. [5] LoCoMo (Maharana et al., ACL 2024). https://arxiv.org/abs/2402.17753 https://github.com/snap-research/locomo on PithPith's reviewarXiv
  6. [6] CacheBlend. https://arxiv.org/abs/2405.16444 on PithPith's reviewarXiv
  7. [7] EPIC. https://arxiv.org/abs/2410.15332 on PithPith's reviewarXiv
  8. [8] TurboRAG. https://arxiv.org/abs/2410.07590 on PithPith's reviewarXiv
  9. [9] Prompt Cache. https://arxiv.org/abs/2311.04934 on PithPith's reviewarXiv
  10. [10] DeepSeek-V2 (MLA). https://arxiv.org/abs/2405.04434 on PithPith's reviewarXiv
  11. [11] Candidate model checkpoints. Config-verified at pinned revision (con- fig.json re-fetched at the SHA, retrieved 2026-09-27): MiniCPM3-4B https://huggingface.co/openbmb/MiniCPM3-4B @ d6b14dda; DeepSeek- V2-Lite https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite @ 604d5664; Qwe... on Pith
  12. [12] KIVI https://arxiv.org/abs/2402.02750; KVQuant https://arxiv.org/abs/ 2401.18079 on PithPith's reviewarXiv
  13. [13] SCBench. https://arxiv.org/abs/2412.10319 on PithPith's reviewarXiv
  14. [14] Pitfalls of KV Cache Compression. https://arxiv.org/abs/2510.00231 on PithPith's reviewarXiv
  15. [15] Kimi Linear . https://arxiv.org/abs/2510.26692 on PithPith's reviewarXiv
  16. [16] MemGPT / Letta. https://arxiv.org/abs/2310.08560 on PithPith's reviewarXiv
  17. [17] MemoryBank. https://arxiv.org/abs/2305.10250 on PithPith's reviewarXiv
  18. [18] MemoryOS. https://arxiv.org/abs/2506.06326 on PithPith's reviewarXiv
  19. [19] OpenViking. https://docs.openviking.ai Vendor documentation on Pith
  20. [20] LangMem. https://github.com/langchain-ai/langmem @ 9d033b47. Vendor source (LangChain), pinned on Pith
show 37 more
  1. [21] Cognee. https://cognee.ai Vendor documentation. 43 on Pith
  2. [22] CAG ”Don’t Do RAG”. https://arxiv.org/abs/2412.15605 on PithPith's reviewarXiv
  3. [23] vLLM PagedAttention. https://arxiv.org/abs/2309.06180 on PithPith's reviewarXiv
  4. [24] SGLang RadixAttention. https://arxiv.org/abs/2312.07104 on PithPith's reviewarXiv
  5. [25] CacheGen. https://arxiv.org/abs/2310.07240 on PithPith's reviewarXiv
  6. [26] LMCache. https://arxiv.org/abs/2510.09665 on PitharXiv
  7. [27] Mooncake. https://arxiv.org/abs/2407.00079 on PithPith's reviewarXiv
  8. [28] MemServe. https://arxiv.org/abs/2406.17565 on PithPith's reviewarXiv
  9. [29] Tutti. https://arxiv.org/abs/2605.03375 on PithPith's reviewarXiv
  10. [30] Memorizing Transformers. https://arxiv.org/abs/2203.08913 on PithPith's reviewarXiv
  11. [31] KV-cache management survey . https://arxiv.org/abs/2607.02574 on PithPith's reviewarXiv
  12. [32] ”Fidelity Before Structure: Verbatim Chunks Beat Lossy Artifact Extraction in Long-Conversation LLM Memory” (T . An). https://arxiv.org/abs/2601. 00821v4 Preprint. Single author , under review on Pith
  13. [33] Knowledge Objects for Persistent LLM Memory . https://arxiv.org/abs/2603. 17781 Preprint on Pith
  14. [34] ConvoMem. https://arxiv.org/abs/2511.10523 Vendor-authored preprint (Salesforce) on PitharXiv
  15. [35] Mehl et al. 2007 (Science). https://pubmed.ncbi.nlm.nih.gov/17615349 Con- text only (order-of-magnitude anchor); the article’s scenarios are declared as- sumptions on Pith
  16. [36] Tokens-per-word planning ratio. Declared assumption (labelled parameter , mea- sured on the chosen tokenizer) on Pith
  17. [37] Meta Zstandard compression benchmark docs. https://github.com/facebook/ zstd/blob/dev/programs/README.md on Pith
  18. [38] BEAM. https://arxiv.org/abs/2510.27246 on PitharXiv
  19. [39] RULER (Hsieh et al., COLM 2024, NVIDIA). https://arxiv.org/abs/2404. 06654 https://github.com/NVIDIA/RULER on Pith
  20. [40] NoLiMa. https://arxiv.org/abs/2502.05167 on PithPith's reviewarXiv
  21. [41] PersonaMem. https://arxiv.org/abs/2504.14225 https://github.com/ bowen-upenn/PersonaMem on PitharXiv
  22. [42] MemoryAgentBench. https://arxiv.org/abs/2507.05257 on PithPith's reviewarXiv
  23. [43] LoCoMo-Refined. https://github.com/mem-eval-suite/LoCoMo_refined Non- archival (repository) on Pith
  24. [44] PerLTQA. https://arxiv.org/abs/2402.16288 44 on PithPith's reviewarXiv
  25. [45] PrefEval. https://arxiv.org/abs/2502.09597 on PithPith's reviewarXiv
  26. [46] MemBench. https://arxiv.org/abs/2506.21605 on PithPith's reviewarXiv
  27. [47] HaluMem. https://arxiv.org/abs/2511.03506 https://github.com/ MemTensor/HaluMem on PitharXiv
  28. [48] MSC / DMR. https://arxiv.org/abs/2107.07567 on PithPith's reviewarXiv
  29. [49] LongMemEval-V2. https://arxiv.org/abs/2605.12493 on PithPith's reviewarXiv
  30. [50] MemoryArena. https://arxiv.org/abs/2602.16313 on PithPith's reviewarXiv
  31. [51] Memora (”From Recall to Forgetting”, Uddin et al.). https://arxiv.org/abs/ 2604.20006 on PithPith's reviewarXiv
  32. [52] Lost in the Middle (Liu et al., TACL 2023). https://arxiv.org/abs/2307.03172 on PithPith's reviewarXiv
  33. [53] GDPR Art. 17. https://eur-lex.europa.eu/legal-content/EN/ALL/?uri= celex:32016R0679 on Pith
  34. [54] KVEraser .https://arxiv.org/abs/2606.17034 on PithPith's reviewarXiv
  35. [55] HELMET (Yen et al., ICLR 2025). https://arxiv.org/abs/2410.02694 https: //github.com/princeton-nlp/HELMET on PithPith's reviewarXiv
  36. [56] Same Task, More Tokens (Levy , Jacoby & Goldberg, ACL 2024). https: //aclanthology.org/2024.acl-long.818/ https://arxiv.org/abs/2402. 14848 on Pith
  37. [57] NVIDIA HGX AI Factory reference architecture, components (H200/B200/B300 SXM memory per GPU). https://docs.nvidia. com/enterprise-reference-architectures/hgx-ai-factory/latest/ components.html; H200 product page https://www.nvidia.com/en-us/ data-center/h200/; DGX B300 user gu... on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment