pith. sign in

arxiv: 2606.26806 · v1 · pith:WH2PPNWMnew · submitted 2026-06-25 · 💻 cs.AI · cs.LG

Memory Depth, Not Memory Access: Selective Parametric Consolidation for Long-Running Language Agents

Haoliang Han This is my paper

Pith reviewed 2026-06-26 04:51 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords memory depthparametric consolidationLoRAlanguage agentsretrievalgoal persistenceloop-drift protocol
0
0 comments X

The pith

Selective parametric consolidation supplies memory depth in language agents distinct from retrieval access.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that long-running language agents require memory depth—durable goal-conditioned tendencies written into a small parametric store—separate from retrieval systems that only fetch facts at query time. Retrieval handles shallow factual recall effectively, but does not decide which experiences should shape future behavior after the working context is unloaded. The loop-drift protocol isolates this distinction by keeping the retrieval index intact while unloading context and testing persistence under long-loop interference. Selective consolidation via EVAF, using surprise- and valence-gated LoRA updates, achieves strong goal persistence with minimal writes, showing the two memory mechanisms are complementary.

Core claim

Selective parametric consolidation supplies memory depth distinct from and complementary to retrieval access. EVAF achieves goal persistence and post-unload recovery scores of 0.812-0.904 across GPT-2 and TinyLlama with only 2-3 parametric writes per 200 events, while retrieval leads on shallow factual recall at 0.956-0.973. The mechanism factorizes into controllable selection and actuation dimensions, with model-dependent inner-loop write strength and asymmetric coupling under miscalibration.

What carries the argument

EVAF, the surprise- and valence-gated LoRA consolidation mechanism that performs selective parametric writes to create durable goal-conditioned behavior.

If this is right

  • Retrieval is strongest on shallow factual recall while EVAF is strongest on goal persistence and post-unload recovery.
  • Selective consolidation requires only 2-3 parametric writes per 200 events.
  • Selection and actuation factorize into two controllable dimensions.
  • Inner-loop write strength is model-dependent across GPT-2, TinyLlama, and Mistral-7B.
  • A matched-gate inversion on Mistral-7B reveals asymmetric selection-actuation coupling under miscalibrated actuation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This separation suggests long-running agents could maintain persistent goals through sparse updates without retaining full context or constant retrieval calls.
  • Public Memora event streams point to stale-memory invalidation as an unresolved limit for any parametric store.
  • Testing the protocol on larger models could clarify how selection-actuation coupling scales.

Load-bearing premise

The loop-drift protocol successfully isolates memory depth by keeping the retrieval index intact while unloading working context and requiring goal-conditioned behavior to persist under long-loop interference.

What would settle it

If goal-conditioned behavior fails to persist after context unload in loop-drift tests when using EVAF but succeeds with retrieval alone, or if selective gates show no advantage over random sparse writes, the distinction between memory depth and access collapses.

Figures

Figures reproduced from arXiv: 2606.26806 by Haoliang Han.

Figure 1
Figure 1. Figure 1: The expected winner flips with memory depth. Retrieval is strongest on shallow factual access, while [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
read the original abstract

Long-running language agents need more than memory access. Retrieval systems can fetch past facts at query time, but they do not decide which experiences should continue to shape behavior after the working context is unloaded. We study this separate problem as memory depth: durable goal-conditioned tendencies written into a small parametric store. We introduce the loop-drift protocol, a controlled stress test in which the retrieval index remains intact while working context is unloaded and goal-conditioned behavior must persist under long-loop interference. We evaluate EVAF, a surprise- and valence-gated LoRA consolidation mechanism. Across GPT-2 and TinyLlama, retrieval is strongest on shallow factual recall (short-fact accuracy 0.956--0.973), while EVAF is strongest on goal persistence and post-unload recovery (0.812--0.904) with only 2--3 parametric writes per 200 events. Mechanism controls show that selective consolidation factorizes into two controllable dimensions: selection and actuation. Matched random gates isolate selection beyond sparse writing; fixed-inner controls across GPT-2, TinyLlama, and Mistral-7B show that inner-loop write strength is model-dependent; and a Mistral-7B matched-gate inversion reveals asymmetric selection-actuation coupling under miscalibrated actuation. Public Memora event streams serve as an external diagnostic, exposing stale-memory invalidation as an unresolved boundary. Within this probe, selective parametric consolidation supplies memory depth distinct from and complementary to retrieval access.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper claims that selective parametric consolidation via a gated LoRA mechanism (EVAF) provides 'memory depth'—durable goal-conditioned tendencies in a parametric store—that is distinct from and complementary to retrieval access. This is demonstrated using the loop-drift protocol, which unloads working context while keeping the retrieval index intact and tests persistence under long-loop interference. Experiments on GPT-2, TinyLlama, and Mistral-7B show retrieval achieving 0.956-0.973 on short-fact accuracy while EVAF achieves 0.812-0.904 on goal persistence with 2-3 writes per 200 events. Mechanism controls factorize selection and actuation, and Memora streams diagnose stale memory issues.

Significance. If the results hold under scrutiny, the work establishes a practical distinction between memory access and memory depth in long-running agents, offering an efficient parametric approach for maintaining goal-directed behavior post-context unload. The use of controllable mechanism ablations (random gates, fixed-inner, matched-gate inversion) and external public data streams provides strong support for the factorization claim and highlights boundaries like stale-memory invalidation. This could influence agent architectures by reducing dependence on retrieval for persistence.

major comments (1)
  1. [Experimental evaluation (referenced via abstract metrics)] The reported performance metrics (e.g., goal persistence 0.812-0.904, short-fact accuracy 0.956-0.973) lack accompanying details on the loop-drift protocol implementation, choice of baselines, number of trials, error bars, statistical significance, or data exclusion rules. This makes it impossible to evaluate whether the protocol successfully isolates memory depth as claimed, directly impacting the soundness of the central claim.
minor comments (2)
  1. Clarify the expansion or origin of the acronym EVAF if it is not a standard term.
  2. [Discussion] The unresolved boundary on stale-memory invalidation is noted but could be expanded with suggestions for future work to strengthen the paper's completeness.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for highlighting the need for greater experimental transparency. We address the single major comment below and will incorporate the requested details in revision.

read point-by-point responses

  1. Referee: [Experimental evaluation (referenced via abstract metrics)] The reported performance metrics (e.g., goal persistence 0.812-0.904, short-fact accuracy 0.956-0.973) lack accompanying details on the loop-drift protocol implementation, choice of baselines, number of trials, error bars, statistical significance, or data exclusion rules. This makes it impossible to evaluate whether the protocol successfully isolates memory depth as claimed, directly impacting the soundness of the central claim.

    Authors: We agree that the manuscript as submitted does not supply sufficient implementation-level detail for independent evaluation of the loop-drift protocol or the reported metrics. In the revised version we will add a dedicated experimental-methods subsection that (1) fully specifies the loop-drift protocol (context-unload schedule, interference length, retrieval-index preservation rules), (2) lists all baselines and controls with their exact configurations, (3) reports the number of independent trials per condition, (4) includes error bars and the statistical tests used, and (5) states any data-exclusion criteria. These additions will allow readers to assess whether the protocol isolates memory depth as intended. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper introduces the loop-drift protocol as a new experimental stress test and evaluates the EVAF mechanism through controlled experiments reporting differential metrics (e.g., retrieval accuracy 0.956-0.973 vs. EVAF goal persistence 0.812-0.904). No equations, derivations, fitted parameters renamed as predictions, or self-citations appear in the provided text. The central claim of distinct memory depth rests on empirical factorization via mechanism controls rather than reducing to inputs by construction, rendering the work self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 3 invented entities

Abstract introduces new concepts without referencing prior literature for their definitions or validation; based solely on abstract text.

free parameters (1)
  • parametric writes per events = 2-3
    Reported as 2-3 writes per 200 events as part of EVAF performance.
axioms (1)
  • domain assumption Retrieval systems fetch facts but do not decide which experiences shape behavior after context unload
    Stated as the separation between retrieval and memory depth in the opening sentences.
invented entities (3)
  • memory depth no independent evidence
    purpose: durable goal-conditioned tendencies written into parametric store
    New term defined to distinguish from retrieval access.
  • EVAF no independent evidence
    purpose: surprise- and valence-gated LoRA consolidation mechanism
    New mechanism introduced for selective consolidation.
  • loop-drift protocol no independent evidence
    purpose: controlled stress test isolating memory depth under context unload
    New evaluation protocol described.

pith-pipeline@v0.9.1-grok · 5790 in / 1381 out tokens · 34516 ms · 2026-06-26T04:51:44.560938+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 10 linked inside Pith

  1. [1]

    and McNaughton, Bruce L

    McClelland, James L. and McNaughton, Bruce L. and O'Reilly, Randall C. , title =. Psychological Review , volume =. 1995 , doi =

    1995

  2. [2]

    , title =

    Kumaran, Dharshan and Hassabis, Demis and McClelland, James L. , title =. Trends in Cognitive Sciences , volume =. 2016 , doi =

    2016

  3. [3]

    Neuron , volume =

    Eichenbaum, Howard , title =. Neuron , volume =. 2017 , doi =

    2017

  4. [4]

    and Turk-Browne, Nicholas B

    Schapiro, Anna C. and Turk-Browne, Nicholas B. and Botvinick, Matthew M. and Norman, Kenneth A. , title =. Philosophical Transactions of the Royal Society B , volume =. 2017 , doi =

    2017

  5. [5]

    Nature Reviews Neuroscience , volume =

    Friston, Karl , title =. Nature Reviews Neuroscience , volume =. 2010 , doi =

    2010

  6. [6]

    , title =

    French, Robert M. , title =. Trends in Cognitive Sciences , volume =. 1999 , doi =

    1999

  7. [7]

    and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and others , title =

    Kirkpatrick, James and Pascanu, Razvan and Rabinowitz, Neil and Veness, Joel and Desjardins, Guillaume and Rusu, Andrei A. and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and others , title =. Proceedings of the National Academy of Sciences , volume =. 2017 , doi =

    2017

  8. [8]

    International Conference on Machine Learning (ICML) , year =

    Zenke, Friedemann and Poole, Ben and Ganguli, Surya , title =. International Conference on Machine Learning (ICML) , year =

  9. [9]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Rolnick, David and Ahuja, Arun and Schwarz, Jonathan and Lillicrap, Timothy and Wayne, Greg , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  10. [10]

    and Torr, Philip H

    Chaudhry, Arslan and Rohrbach, Marcus and Elhoseiny, Mohamed and Ajanthan, Thalaiyasingam and Dokania, Puneet K. and Torr, Philip H. S. and Ranzato, Marc'Aurelio , title =. ICML Workshop on Multi-Task and Lifelong Reinforcement Learning , year =

  11. [11]

    and Rabinowitz, Neil C

    Rusu, Andrei A. and Rabinowitz, Neil C. and Desjardins, Guillaume and Soyer, Hubert and Kirkpatrick, James and Kavukcuoglu, Koray and Pascanu, Razvan and Hadsell, Raia , title =. arXiv preprint arXiv:1606.04671 , year =

    Pith/arXiv arXiv

  12. [12]

    European Conference on Computer Vision (ECCV) , year =

    Mallya, Arun and Davis, Dillon and Lazebnik, Svetlana , title =. European Conference on Computer Vision (ECCV) , year =

  13. [13]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =. International Conference on Learning Representations (ICLR) , year =

  14. [14]

    International Conference on Learning Representations (ICLR) , year =

    Razdaibiedina, Anastasia and Mao, Yuning and Hou, Rui and Khabsa, Madian and Lewis, Mike and Almahairi, Amjad , title =. International Conference on Learning Representations (ICLR) , year =

  15. [15]

    arXiv preprint arXiv:2311.02428 , year =

    Chitale, Rajas and Vaidya, Ankit and Kane, Aditya and Ghotkar, Archana , title =. arXiv preprint arXiv:2311.02428 , year =

    arXiv

  16. [16]

    International Conference on Learning Representations (ICLR) , year =

    Kemker, Ronald and Kanan, Christopher , title =. International Conference on Learning Representations (ICLR) , year =

  17. [17]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Javed, Khurram and White, Martha , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  18. [18]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Mujika, Asier and Meier, Florian and Steger, Angelika , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  19. [19]

    International Conference on Machine Learning (ICML) , year =

    Das, Payel and Chaudhury, Subhajit and Nelson, Elliot and Melnyk, Igor and Swaminathan, Sarath and Daheim, Nico and Lozano, Aurelie and Ross, Brian and Tabaja, Niloofar and Gandhi, Akshay and others , title =. International Conference on Machine Learning (ICML) , year =

  20. [20]

    Retrieval-augmented generation for knowledge-intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-augmented generation for knowledge-intensive. Advances in Neural Information Processing Systems (NeurIPS) , year =

  21. [21]

    Journal of Machine Learning Research , volume =

    Izacard, Gautier and Lewis, Patrick and Lomeli, Maria and Hosseini, Lucas and Petroni, Fabio and Schick, Timo and Dwivedi-Yu, Jane and Joulin, Armand and Riedel, Sebastian and Grave, Edouard , title =. Journal of Machine Learning Research , volume =

  22. [22]

    and Hutchins, DeLesley and Szegedy, Christian , title =

    Wu, Yuhuai and Rabe, Markus N. and Hutchins, DeLesley and Szegedy, Christian , title =. International Conference on Learning Representations (ICLR) , year =

  23. [23]

    and Stoica, Ion and Gonzalez, Joseph E

    Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , title =. arXiv preprint arXiv:2310.08560 , year =

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2405.14831 , year =

    Guti. arXiv preprint arXiv:2405.14831 , year =

    arXiv

  25. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Behrouz, Ali and Zhong, Peilin and Mirrokni, Vahab , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  26. [26]

    International Conference on Learning Representations (ICLR) , year =

    Feng, Guhao and Luo, Shengjie and Hua, Kai and Zhang, Ge and Huang, Wenhao and He, Di and Cai, Tianle , title =. International Conference on Learning Representations (ICLR) , year =

  27. [27]

    arXiv preprint arXiv:2606.04536 , year =

    Ren, Tao and Luo, Weiyao and Yang, Hui and Zhu, Rongzhi and Huang, Xiang and Wu, Yuchuan and Chou, Bingxue and Ye, Jieping and Liang, Jiafeng and Li, Yongbin and Peng, Yijie , title =. arXiv preprint arXiv:2606.04536 , year =

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2606.05698 , year =

    Zuo, Chunsheng and Wang, Liaoyaqi and Jurayj, William and Fleshman, William and Van Durme, Benjamin , title =. arXiv preprint arXiv:2606.05698 , year =

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2605.30260 , year =

    Xu, Ziwen and Hong, Haiwen and Yu, Linsong and Cui, Benglei and Huang, Longtao and Xue, Hui and Zhang, Ningyu , title =. arXiv preprint arXiv:2605.30260 , year =

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2606.22844 , year =

    Yang, Wei and Kan, Bryce and Li, Shixuan and Li, Li and Qin, Yuehan and Li, Jiate and Bogdan, Paul and Thomason, Jesse , title =. arXiv preprint arXiv:2606.22844 , year =

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2606.19172 , year =

    Li, Bojie , title =. arXiv preprint arXiv:2606.19172 , year =

    Pith/arXiv arXiv

  32. [32]

    Ross , title =

    Tavakoli, Mohammad and Salemi, Alireza and Ye, Carrie and Abdalla, Mohamed and Zamani, Hamed and Mitchell, J. Ross , title =. arXiv preprint arXiv:2510.27246 , year =

    arXiv

  33. [33]

    arXiv preprint arXiv:2501.13956 , year =

    Rasmussen, Preston and Paliychuk, Pavlo and Beauvais, Travis and Ryan, Jack and Chalef, Daniel , title =. arXiv preprint arXiv:2501.13956 , year =

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2410.10813 , year =

    Wu, Di and Wang, Hongwei and Yu, Wenhao and Zhang, Yunsheng and Chang, Kai-Wei and Yu, Dong , title =. arXiv preprint arXiv:2410.10813 , year =

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2402.17753 , year =

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , title =. arXiv preprint arXiv:2402.17753 , year =

    Pith/arXiv arXiv

  36. [36]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  37. [37]

    Anthropic transformer-circuits.pub , year =

    Elhage, Nelson and Hume, Tristan and Olsson, Catherine and Schiefer, Nicholas and Henighan, Tom and Kravec, Shauna and Hatfield-Dodds, Zac and Lasenby, Robert and Drain, Dawn and Chen, Carol and others , title =. Anthropic transformer-circuits.pub , year =

  38. [38]

    International Conference on Machine Learning (ICML) , year =

    Park, Kiho and Choe, Yo Joong and Veitch, Victor , title =. International Conference on Machine Learning (ICML) , year =

  39. [39]

    EMNLP , year =

    Geva, Mor and Caciularu, Avi and Wang, Kevin Ro and Goldberg, Yoav , title =. EMNLP , year =

  40. [40]

    International Conference on Learning Representations (ICLR) , year =

    Gurnee, Wes and Tegmark, Max , title =. International Conference on Learning Representations (ICLR) , year =

  41. [41]

    International Conference on Learning Representations (ICLR) , year =

    Hernandez, Evan and Sharma, Arnab Sen and Haklay, Tal and Meng, Kevin and Wattenberg, Martin and Andreas, Jacob and Belinkov, Yonatan and Bau, David , title =. International Conference on Learning Representations (ICLR) , year =

  42. [42]

    EMNLP , year =

    De Cao, Nicola and Aziz, Wilker and Titov, Ivan , title =. EMNLP , year =

  43. [43]

    arXiv preprint arXiv:2211.11031 , year =

    Hartvigsen, Thomas and Sankaranarayanan, Swami and Palangi, Hamid and Kim, Yoon and Ghassemi, Marzyeh , title =. arXiv preprint arXiv:2211.11031 , year =

    arXiv

  44. [44]

    International Conference on Learning Representations (ICLR) , year =

    Ge, Tao and Hu, Jing and Wang, Lei and Wang, Xun and Chen, Si-Qing and Wei, Furu , title =. International Conference on Learning Representations (ICLR) , year =

  45. [45]

    EMNLP , year =

    Chevalier, Alexis and Wettig, Alexander and Ajith, Anirudh and Chen, Danqi , title =. EMNLP , year =

  46. [46]

    and Cowan, Jack D

    Wilson, Hugh R. and Cowan, Jack D. , title =. Biophysical Journal , volume =. 1972 , doi =

    1972

  47. [47]

    , title =

    Burak, Yoram and Fiete, Ila R. , title =. PLoS Computational Biology , volume =. 2009 , doi =

    2009

  48. [48]

    Vershynin, Roman , title =

  49. [49]

    ACL , year =

    Zhang, Saizheng and Dinan, Emily and Urbanek, Jack and Szlam, Arthur and Kiela, Douwe and Weston, Jason , title =. ACL , year =