arxiv: 2604.04514 · v1 · submitted 2026-04-06 · 💻 cs.AI · cs.CL· cs.IR

Recognition: 3 theorem links

· Lean Theorem

SuperLocalMemory V3.3: The Living Brain -- Biologically-Inspired Forgetting, Cognitive Quantization, and Multi-Channel Retrieval for Zero-LLM Agent Memory Systems

Varun Pratap Bhardwaj

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:42 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.IR

keywords agent memorycognitive forgettingmulti-channel retrievallocal AI systemsEbbinghaus curvequantization-aware distancezero-LLM memory

0 comments

The pith

A local-first memory system using Ebbinghaus forgetting, FRQAD metric, and seven retrieval channels reaches 70.4 percent on agent conversation benchmarks without any LLM.

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

The paper presents SuperLocalMemory V3.3 as a memory architecture for AI coding agents that incorporates biologically modeled cognitive processes instead of relying on cloud-based LLMs for recall. It claims that combining Fisher-Rao Quantization-Aware Distance for embedding fidelity, adaptive forgetting curves tied to lifecycle compression, and parallel retrieval across semantic, temporal, entity, and associative channels produces measurable gains on multi-hop and adversarial tasks. A sympathetic reader would care because current agent systems store vast parametric knowledge yet lose context from recent exchanges, and a purely local solution that automates memory lifecycle could reduce latency and privacy risks while preserving performance. The work builds directly on prior information-geometric foundations to show that these mechanisms together close much of the gap to human-like persistence in zero-LLM settings.

Core claim

V3.3 implements the full cognitive memory taxonomy through FRQAD as a metric on the Gaussian manifold that prefers high-fidelity embeddings at 100 percent precision, Ebbinghaus Adaptive Forgetting coupled to progressive quantization that yields 6.7 times greater discrimination, and a seven-channel retrieval system spanning semantic, keyword, entity graph, temporal, spreading activation, consolidation, and Hopfield associative pathways, delivering 70.4 percent accuracy on LoCoMo in zero-LLM Mode A along with gains of 23.8 points on multi-hop and 12.7 points on adversarial subsets.

What carries the argument

The seven-channel cognitive retrieval system combined with the FRQAD metric and Ebbinghaus Adaptive Forgetting curve that links forgetting rate to embedding compression over memory lifecycle stages.

If this is right

Agent memory can run entirely on CPU with no cloud LLM calls for core recall operations while still handling complex conversation tasks.
Memory parameterization through soft prompts enables long-term implicit storage that persists across sessions without explicit storage overhead.
The auto-cognitive pipeline automates the complete memory lifecycle from encoding through consolidation and forgetting, reducing manual tuning.
Deliberate trade-offs between Mode A zero-LLM accuracy and higher modes that use LLMs allow designers to choose operating points based on resource constraints.

Where Pith is reading between the lines

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

If the forgetting curve and channel combination prove robust, they could be ported to other local vector stores to improve discrimination without changing embedding models.
The approach suggests that explicit modeling of memory consolidation stages may reduce the need for ever-larger context windows in downstream models.
Success on adversarial subsets implies the multi-channel design could help agents resist prompt-injection style attacks that target single retrieval paths.

Load-bearing premise

That the seven-channel retrieval, FRQAD metric, and lifecycle-aware quantization deliver genuine cognitive advantages that generalize beyond the LoCoMo benchmark and the specific implementation choices in V3.3.

What would settle it

Measure whether the same performance deltas appear when the system is tested on a different long-context agent benchmark that emphasizes multi-turn reasoning outside the LoCoMo distribution while holding all other variables fixed.

Figures

Figures reproduced from arXiv: 2604.04514 by Varun Pratap Bhardwaj.

**Figure 2.** Figure 2: Mixed-precision preference: percentage of 18,840 query-fact pairs where the f32 [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

**Figure 3.** Figure 3: Ebbinghaus retention curves over 30 simulated days. Hot facts (daily access) converge [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

**Figure 4.** Figure 4: LoCoMo per-category comparison. V3.3 R3 surpasses Paper 2 on adversarial (+6.1pp) [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

read the original abstract

AI coding agents operate in a paradox: they possess vast parametric knowledge yet cannot remember a conversation from an hour ago. Existing memory systems store text in vector databases with single-channel retrieval, require cloud LLMs for core operations, and implement none of the cognitive processes that make human memory effective. We present SuperLocalMemory V3.3 ("The Living Brain"), a local-first agent memory system implementing the full cognitive memory taxonomy with mathematical lifecycle dynamics. Building on the information-geometric foundations of V3.2 (arXiv:2603.14588), we introduce five contributions: (1) Fisher-Rao Quantization-Aware Distance (FRQAD) -- a new metric on the Gaussian statistical manifold achieving 100% precision at preferring high-fidelity embeddings over quantized ones (vs 85.6% for cosine), with zero prior art; (2) Ebbinghaus Adaptive Forgetting with lifecycle-aware quantization -- the first mathematical forgetting curve in local agent memory coupled to progressive embedding compression, achieving 6.7x discriminative power; (3) 7-channel cognitive retrieval spanning semantic, keyword, entity graph, temporal, spreading activation, consolidation, and Hopfield associative channels, achieving 70.4% on LoCoMo in zero-LLM Mode A; (4) memory parameterization implementing Long-Term Implicit memory via soft prompts; (5) zero-friction auto-cognitive pipeline automating the complete memory lifecycle. On LoCoMo, V3.3 achieves 70.4% in Mode A (zero-LLM), with +23.8pp on multi-hop and +12.7pp on adversarial. V3.2 achieved 74.8% Mode A and 87.7% Mode C; the 4.4pp gap reflects a deliberate architectural trade-off. SLM V3.3 is open source under the Elastic License 2.0, runs entirely on CPU, with over 5,000 monthly downloads.

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.

Desk Editor's Note private letter to a colleague

V3.3 layers a new Fisher-Rao metric, Ebbinghaus forgetting curve, and seven retrieval channels onto local agent memory, but the LoCoMo gains aren't isolated to those additions.

read the letter

The main thing to know about this paper is that it extends the author's earlier V3.2 work with a new distance metric based on Fisher-Rao geometry, an Ebbinghaus-inspired forgetting curve linked to quantization, and a multi-channel retrieval setup, all aimed at local agent memory. It gets decent results on the LoCoMo benchmark in zero-LLM mode. The paper does a good job describing a complete local system that tries to mimic cognitive processes like forgetting and associative retrieval. Running entirely on CPU and being open source makes it accessible. The claims about 100% precision for the new metric and 6.7x power for forgetting are specific and could be checked if the code is available. Where it gets soft is the lack of evidence tying the improvements directly to these new parts. There are no ablations or controls for the embedding model or parameters, so the attribution to FRQAD or the seven channels remains unproven. Relying on one benchmark also limits how much we can say about broader usefulness. The circularity with the prior version is minor since it's an extension, but independent validation would help. This is for engineers and researchers focused on agent architectures and memory management in AI systems. Someone looking for practical local alternatives to cloud-based memory could find value in the design choices. It deserves a serious referee because the work is grounded in implementation and has clear, testable claims even if the current evaluation is limited. I recommend putting it through peer review with feedback on adding component-wise experiments and testing on additional tasks.

Referee Report

3 major / 2 minor

Summary. The paper presents SuperLocalMemory V3.3 ('The Living Brain'), a local-first, zero-LLM agent memory system that implements a full cognitive memory taxonomy. Building on the authors' prior V3.2 work, it introduces Fisher-Rao Quantization-Aware Distance (FRQAD) as a metric on the Gaussian manifold, Ebbinghaus Adaptive Forgetting coupled to lifecycle-aware quantization, a 7-channel retrieval architecture (semantic, keyword, entity graph, temporal, spreading activation, consolidation, Hopfield), and memory parameterization via soft prompts. The central empirical claims are 70.4% accuracy on LoCoMo in Mode A (zero-LLM), with +23.8pp gains on multi-hop and +12.7pp on adversarial subsets, FRQAD achieving 100% precision at preferring high-fidelity embeddings (vs. 85.6% for cosine), and Ebbinghaus forgetting delivering 6.7x discriminative power.

Significance. If the reported gains can be rigorously attributed to the proposed cognitive mechanisms rather than implementation details, the work could meaningfully advance local, biologically-plausible memory architectures for autonomous agents. The open-source release under Elastic License 2.0, CPU-only execution, and explicit mathematical framing of forgetting and quantization are positive elements that could support reproducibility and extension. However, the absence of isolating experiments currently limits the assessed significance to incremental engineering contributions rather than a validated cognitive advance.

major comments (3)

[Abstract] Abstract and results: The manuscript reports 70.4% LoCoMo Mode A accuracy together with the +23.8pp multi-hop and +12.7pp adversarial lifts, yet supplies no ablation tables, control conditions, or statistical tests that isolate the contributions of FRQAD, Ebbinghaus Adaptive Forgetting, or the seven retrieval channels from other factors (embedding model choice, quantization schedule, channel weighting). Without such controls the attribution of performance to the new cognitive components cannot be verified.
[Abstract] Abstract: No error bars, run-to-run variance, or data-exclusion criteria are stated for any of the quoted performance figures (70.4%, 100% FRQAD precision, 6.7x discriminative power). The experimental design therefore does not meet standard requirements for reproducible claims in AI systems papers.
[Abstract] Abstract and comparison to V3.2: The 4.4pp drop relative to V3.2's 74.8% Mode A score is described as a deliberate architectural trade-off, but no quantitative characterization of that trade-off or results on any benchmark other than LoCoMo are provided. This leaves open whether the V3.3 mechanisms generalize or merely reflect benchmark-specific tuning.

minor comments (2)

[Abstract] Abstract: The claim of 'zero prior art' for FRQAD would be strengthened by a short literature pointer or explicit statement that no comparable information-geometric metric on quantized embeddings has been proposed.
[Abstract] The abstract states that the system 'implements the full cognitive memory taxonomy' but does not define the taxonomy or map each of the seven channels to established cognitive psychology constructs; a brief clarifying sentence would improve accessibility.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and describe the revisions planned to improve experimental rigor and transparency.

read point-by-point responses

Referee: [Abstract] Abstract and results: The manuscript reports 70.4% LoCoMo Mode A accuracy together with the +23.8pp multi-hop and +12.7pp adversarial lifts, yet supplies no ablation tables, control conditions, or statistical tests that isolate the contributions of FRQAD, Ebbinghaus Adaptive Forgetting, or the seven retrieval channels from other factors (embedding model choice, quantization schedule, channel weighting). Without such controls the attribution of performance to the new cognitive components cannot be verified.

Authors: We agree that ablation studies are required to isolate the contributions of the proposed mechanisms. In the revised manuscript we will add ablation tables that disable FRQAD (replacing it with cosine), disable Ebbinghaus forgetting, and ablate individual retrieval channels, reporting the resulting LoCoMo accuracy changes together with paired statistical tests. revision: yes
Referee: [Abstract] Abstract: No error bars, run-to-run variance, or data-exclusion criteria are stated for any of the quoted performance figures (70.4%, 100% FRQAD precision, 6.7x discriminative power). The experimental design therefore does not meet standard requirements for reproducible claims in AI systems papers.

Authors: We acknowledge this omission. The revised version will report standard deviations across multiple runs for all key metrics, specify the number of runs and random seeds, and explicitly document data-exclusion criteria and preprocessing steps for the LoCoMo evaluation. revision: yes
Referee: [Abstract] Abstract and comparison to V3.2: The 4.4pp drop relative to V3.2's 74.8% Mode A score is described as a deliberate architectural trade-off, but no quantitative characterization of that trade-off or results on any benchmark other than LoCoMo are provided. This leaves open whether the V3.3 mechanisms generalize or merely reflect benchmark-specific tuning.

Authors: We will expand the manuscript to quantify the trade-off with concrete metrics on latency, memory footprint, and CPU usage that motivated the zero-LLM design. We do not currently have results on other benchmarks, as evaluation was focused on LoCoMo; we will add a limitations discussion on generalization and note this as future work. revision: partial

standing simulated objections not resolved

Results on benchmarks other than LoCoMo, which were not performed in the present study.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's abstract describes a memory system with new components (FRQAD metric, Ebbinghaus Adaptive Forgetting, 7-channel retrieval) and reports empirical results on the LoCoMo benchmark (70.4% Mode A score, specific percentage point lifts). It references prior same-author work for foundations but presents the performance numbers as measured outcomes rather than as outputs of a closed mathematical derivation. No equations, fitted parameters renamed as predictions, or self-referential definitions are visible in the provided text that would reduce the central claims to their inputs by construction. The self-citation supplies context but does not bear the load of the benchmark results, which are independent of the prior paper. This is a standard non-circular empirical system paper.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility into parameters; the design assumes cognitive taxonomy can be directly mapped to mathematical dynamics and retrieval channels without showing independent evidence for those mappings.

free parameters (2)

Quantization compression schedule
Progressive embedding compression tied to forgetting curve is required for the 6.7x claim but no values or fitting procedure given.
Channel weighting parameters
Seven retrieval channels must be combined; weights or fusion rules are not specified.

axioms (1)

domain assumption Human cognitive memory taxonomy can be faithfully implemented via statistical manifold metrics and lifecycle dynamics in software
Invoked to justify the full taxonomy and mathematical forgetting curve.

pith-pipeline@v0.9.0 · 5686 in / 1415 out tokens · 85633 ms · 2026-05-10T19:42:52.630868+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

FRQAD ... dFR = sqrt( sum [sqrt(2) arccosh(1 + (μ1k-μ2k)^2 + 2(σ1k-σ2k)^2)/(4σ1kσ2k)]^2 ) ... on the Gaussian statistical manifold
IndisputableMonolith/Foundation/ArrowOfTime.lean arrow_from_z unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Ebbinghaus Adaptive Forgetting ... R(m,t) = exp(-t/S(m)) ... coupled to Fokker-Planck lifecycle dynamics ... Theorem 5.3 (Convergence)
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

7-channel cognitive retrieval ... weighted Reciprocal Rank Fusion ... Hopfield associative memory

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 13 canonical work pages · 2 internal anchors

[1]

MemoryLLM: Towards self-updatable large language models. 2024

2024
[2]

Context-as-memory.arXiv preprint arXiv:2506.03141, 2025

work page arXiv 2025
[3]

When less is more: 8-bit quantization improves continual learning in LLMs.arXiv preprint arXiv:2512.18934, 2025

work page arXiv 2025
[4]

FOREVER: Forgetting curve-inspired memory replay for continual learning.arXiv preprint arXiv:2601.03938, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[5]

Natural gradient works efficiently in learning.Neural Computation, 10(2): 251–276, 1998

Shun-ichi Amari. Natural gradient works efficiently in learning.Neural Computation, 10(2): 251–276, 1998

1998
[6]

Colin Atkinson and Ann F. S. Mitchell. Rao’s distance measure.Sankhy¯ a: The Indian Journal of Statistics, Series A, 43(3):345–365, 1981

1981
[7]

Agent Behavioral Contracts: Formal Specification and Runtime Enforcement,

Varun Pratap Bhardwaj. Privacy-preserving multi-agent memory with Bayesian trust defense.arXiv preprint arXiv:2602.22302, 2026

work page arXiv 2026
[8]

SuperLocalMemory v3: Information-geometric cognitive memory for AI agents.arXiv preprint arXiv:2603.14588, 2026

Varun Pratap Bhardwaj. Information-geometric foundations for zero-LLM enterprise agent memory.arXiv preprint arXiv:2603.14588, 2026

work page arXiv 2026
[9]

Collins and Elizabeth F

Allan M. Collins and Elizabeth F. Loftus. A spreading-activation theory of semantic processing.Psychological Review, 82(6):407–428, 1975

1975
[10]

Duncker & Humblot, Leipzig, 1885

Hermann Ebbinghaus.Über das Gedächtnis. Duncker & Humblot, Leipzig, 1885
[11]

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. 2024

2024
[12]

PolarQuant: Polar-Coordinate KV Cache Quantization,

Insu Han, Praneeth Kacham, Amin Karbasi, Vahab Mirrokni, and Amir Zandieh. Polar- Quant: Quantizing KV caches with polar transformation. InProceedings of AISTATS, 2026. arXiv:2502.02617

work page arXiv 2026
[13]

SYNAPSE: Synergistic associative processing & semantic encoding

Hanqi Jiang et al. SYNAPSE: Synergistic associative processing & semantic encoding. arXiv preprint arXiv:2601.02744, 2026

work page arXiv 2026
[14]

Cognitive Memory in Large Language Models, April 2025

Zhongyang Li et al. Cognitive memory in large language models.arXiv preprint arXiv:2504.02441, 2025

work page arXiv 2025
[15]

Evaluating very long-term conversational memory of LLM agents

Priyanka Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of LLM agents. In Proceedings of ACL, 2024. arXiv:2402.09714. 18

work page arXiv 2024
[16]

McClelland, Bruce L

James L. McClelland, Bruce L. McNaughton, and Randall C. O’Reilly. Why there are complementary learning systems in the hippocampus and neocortex.Psychological Review, 102(3), 1995

1995
[17]

Mem0: The memory layer for personalized AI.https://github.com/mem0ai/ mem0, 2024

Mem0 AI. Mem0: The memory layer for personalized AI.https://github.com/mem0ai/ mem0, 2024

2024
[18]

How to generate random matrices from the classical compact groups

Francesco Mezzadri. How to generate random matrices from the classical compact groups. Notices of the AMS, 54(5):592–604, 2007

2007
[19]

MEM1: RL-trained memory consolidation for LLM agents, 2025

MIT/NUS. MEM1: RL-trained memory consolidation for LLM agents, 2025

2025
[20]

MemGPT: Towards LLMs as Operating Systems

Charles Packer et al. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[21]

Hopfield networks is all you need.arXiv preprint arXiv:2008.02217, 2020

Hubert Ramsauer et al. Hopfield networks is all you need. InProceedings of ICLR, 2021. arXiv:2008.02217

work page arXiv 2021
[22]

QJL: 1-Bit Quantized JL Transform for KV Cache Quantization with Zero Overhead,

Amir Zandieh, Majid Daliri, and Insu Han. QJL: 1-bit quantized JL transform for KV cache quantization with zero overhead. InProceedings of AAAI, 2025. arXiv:2406.03482

work page arXiv 2025
[23]

Turboquant: Online vector quantization with near-optimal distortion rate,

Amir Zandieh, Majid Daliri, Majid Hadian, and Vahab Mirrokni. TurboQuant: Online vector quantization with near-optimal distortion rate. InProceedings of ICLR, 2026. arXiv:2504.19874. 19

work page arXiv 2026