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

Varun Pratap Bhardwaj

Source-linked AI summary

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

Varun Pratap Bhardwaj

arXiv:2604.04514v1Apr 6, 2026cs.AIcs.CLcs.IR

TL;DR

AI coding agents and existing memory tools do not provide persistent, cognitively organized memory across sessions without substantial cloud dependence. SuperLocalMemory V3.3 implements a local cognitive-memory system with mathematical forgetting, quantization-aware retrieval, seven retrieval channels, soft-prompt parameterization, and automatic lifecycle management. It reports 70.4% LoCoMo accuracy in zero-LLM Mode A, while documenting a cold-start requirement and single-hop regression.

Problem
Existing agent memory systems remain concentrated in explicit external storage and do not implement transitions to implicit memory or mathematical forgetting.
Method
The paper implements the cognitive memory taxonomy through mathematical lifecycle dynamics, quantization-aware distance, multi-channel retrieval, soft prompts, and automatic hooks.
Results
70.4% LoCoMo accuracy is achieved in zero-LLM Mode A, with +23.8pp on multi-hop queries.
Takeaways & Limitations
The system provides a local implementation of the complete cognitive memory taxonomy, with convergent mathematics and production deployment reported by the authors.
Takeaways & Limitations
Cold-start adaptation requires approximately 200 feedback signals, while seven-channel fusion introduces noise on simple queries and causes a LoCoMo single-hop regression.

Abstract

from arXiv · show

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.

1 Introduction

AI coding agents lack persistent cross-session memory, while existing systems largely provide static vector retrieval or cloud-dependent operations. SuperLocalMemory V3.3 addresses this gap with a local cognitive-memory architecture spanning lifecycle management, multi-channel retrieval, parameterization, and automation.

The Session Amnesia Problem: Persistent cross-session memory is needed because context-window information disappears when a session ends, forcing developers to repeatedly restore project knowledge.The paper distinguishes ephemeral context from memory that accumulates and organizes knowledge across a developer’s interactions.
The Cognitive Memory Gap: Existing agent memory systems primarily store text in vector databases and retrieve it by similarity without forgetting, consolidation, compression, or parameterization.The paper contrasts this static-store model with active processes attributed to human memory.
The Cognitive Memory Gap: SuperLocalMemory V3.3 claims to span all four cognitive memory tiers with mathematical foundations for transitions while operating entirely on local hardware.The four tiers include sensory, short-term, long-term explicit, and long-term implicit memory.
Contributions: FRQAD achieves 100% precision at preferring high-fidelity embeddings over quantized ones, compared with 85.6% for cosine similarity, while Local TurboQuant targets persistent memory stores.The contribution combines an information-geometric distance with data-oblivious quantization for persistent embeddings.
Contributions: Ebbinghaus Adaptive Forgetting couples retention dynamics to progressive embedding precision, producing 6.7× discriminative power between frequently accessed and unused memories over 30 simulated days.The lifecycle moves embeddings from 32-bit active storage toward lower-precision states as memories fade.

2 Background and Related Work

Prior systems address agent memory through cloud services, flat retrieval, or isolated cognitive mechanisms, but the paper identifies missing combinations of persistent quantization, forgetting, geometric similarity, and implicit parameterization. SuperLocalMemory V3.3 presents these capabilities within a local SQLite-based architecture.

Agent Memory Systems: Mem0, Letta, and Zep provide memory capabilities but involve cloud dependence or static retrieval limitations, while the paper states that existing systems lack forgetting, consolidation, compression, and parameterization.Reported LoCoMo scores include 64.2% for Mem0 and 85.2% for Zep’s Graphiti engine.
Vector Compression and Quantization: The paper positions SLM’s quantization work as the first application of data-oblivious methods to persistent agent memory stores, where mixed-precision search must operate across sessions.Prior quantization methods are described as targeting ephemeral KV-cache compression.
Research Gaps: The paper identifies two related gaps: no prior system combines vector quantization with information-geometric distance metrics, and no local system couples mathematical forgetting to embedding precision.The claimed forgetting dynamics include provable convergence and lifecycle-dependent precision.
Forgetting and Quantization: Prior work links forgetting curves to memory systems and quantization noise to reduced catastrophic forgetting, motivating SLM’s forgetting–quantization coupling.The paper describes this coupling as empirically supported by the cited regularization finding.
Cognitive Retrieval: Spreading activation addresses contextual isolation, but the paper identifies no prior system combining it with Fisher-Rao similarity scoring.Context-as-Memory is cited as evidence that non-contiguous retrieval can outperform sliding-window recency.
Memory Parameterization: SLM converts consolidated text memories into API-compatible natural-language soft prompts, addressing the stated gap in automatic conversion of local memory stores into prompt templates.The paper contrasts this approach with systems requiring model internals or gradient access.

3 System Architecture

SLM V3.3 is a modular, local-first system organized into interface, engine, and storage layers, with automated memory processing, multi-channel retrieval, and multiple operating modes.

SLM V3.3 comprises 17 packages, 215 source modules, and 60 MCP tools, backed by SQLite with sqlite-vec.
Interface Layer: The interface layer provides MCP, CLI, dashboard, and auto-cognitive hook entry points for interacting with and managing memory.
Engine Layer: The engine coordinates encoding, seven-channel retrieval with weighted RRF and reranking, lifecycle management, and adaptive learning pipelines.
Storage Layer: The storage layer keeps core facts, graphs, embeddings, temporal data, quantized embeddings, soft prompts, forgetting schedules, and learning signals in local SQLite databases.
The code knowledge graph links parsed functions, classes, and imports to related memories, enabling retrieval of callers, dependencies, and architectural context.
70.4% on LoCoMo is reported for zero-LLM Mode A, while cloud-augmented Mode C reaches 87.7%.

4 Fisher-Rao Quantization-Aware Distance and Local TurboQuant (C1)

This section introduces FRQAD for mixed-precision embedding retrieval and extends TurboQuant with lifecycle-aware quantization for persistent agent memory.

FRQAD is a new distance metric that compares embeddings across quantization levels using information geometry.
TurboQuant is adapted from transient KV-cache vectors to persistent agent memory stores with long lifetimes, random-access search, unbounded growth, and irrecoverable corruption risk.
Random orthogonal rotation and Lloyd-Max codebooks enable per-coordinate quantization based on the rotated coordinates’ high-dimensional distribution.
MSE distortion remains within 2.7× of the information-theoretic lower bound.
FRQAD models embeddings as diagonal Gaussians whose variances incorporate quantization noise, then computes a Fisher-Rao geodesic.
In the current release, all embeddings default to float32, so FRQAD and standard Fisher-Rao are identical until mixed-precision storage is enabled.

5 Ebbinghaus Adaptive Forgetting (C2)

SLM V3.3 formalizes forgetting as adaptive memory-strength decay and couples lifecycle states to progressive embedding quantization within an information-geometric dynamical system.

An always-retained memory store risks retrieval degradation, storage bloat, and context pollution as irrelevant and old memories accumulate.
Memory strength combines access count, importance, confirmation count, and emotional salience, with logarithmic access dependence producing diminishing returns.
Retention is mapped to discrete lifecycle states that organize memories by their current persistence level.
The forgetting extension adds a rate λ(m)=1/S(m), a drift toward Archive/Forgotten states, and an effective temperature determined by Fisher confidence.
Theorem 5.3 states that the resulting Ebbinghaus-Fokker-Planck density converges to a unique stationary distribution under confining-potential and positive-temperature conditions.
Forgetting and quantization are unified: as memories fade, quantization increases effective variance and lowers their Fisher-Rao similarity scores.
Bayesian trust scores modulate decay, with zero-trust sources decaying three times faster than fully trusted sources under the stated default sensitivity.

6 Seven-Channel Cognitive Retrieval (C3)

V3.3 combines seven cognitive retrieval channels through weighted fusion, selective cross-channel intersection, and ONNX reranking to support semantic and associative recall.

The seven-channel architecture includes semantic, keyword, entity graph, temporal, spreading activation, consolidation, and Hopfield associative retrieval.
Hopfield Associative Memory: Hopfield retrieval complements other channels by supporting pattern completion from partial cues and associative recall of composite patterns.
For multi-hop queries, V3.3 intersects entity-channel and temporal-channel results before fusion to reduce dilution from independent-channel noise.
An ONNX cross-encoder reranks query–memory pairs and produces relevance scores.

7 Memory Parameterization (C4)

Memory parameterization converts consolidated patterns into natural-language soft prompts that configure agent behavior without explicit retrieval, implementing the Long-Term Implicit memory tier for API-based agents.

Memory parameterization implements the Long-Term Implicit tier, which existing systems do not provide.
In implicit memory, past experience configures behavior without explicit retrieval.
High-confidence patterns from consolidated memories become natural-language templates injected at session start.
The pipeline clusters episodic memories, extracts semantic patterns, and filters them using minimum confidence 0.7 and five observations.
Template-based soft-prompt generation and session-start injection require zero LLM cost and are capped at 1,500 tokens.
Unlike LoRA, SLM soft prompts work with any LLM, API, or provider without model-weight access.

8 Zero-Friction Auto-Cognitive Pipeline (C5)

The zero-friction pipeline automates memory operations from session recall through observation, saving, consolidation, parameterization, forgetting, and renewed recall, without requiring manual invocation.

Existing memory tools require manual invocation, creating an adoption barrier because users must remember to use them.
A single npm install -g superlocalmemory triggers automatic lifecycle management.
Session-start hooks load project context, memories, patterns, and soft prompts, while coding hooks observe file changes with five-minute per-file rate limiting.
At session end, the system generates a session summary with git context.
Between sessions, Ebbinghaus decay, precision scheduling, and consolidation complete the lifecycle through automatic recall, observation, saving, learning, consolidation, parameterization, and forgetting.
All hooks fail silently, avoid PreToolUse gates, and support explicit opt-out through slm hooks remove.

9 Compliance and Trust

The system combines local compliance and trust controls with benchmarks for quantization, memory efficiency, forgetting, continuity, and LoCoMo retrieval, where stronger complex-query performance accompanies a single-hop trade-off.

Compliance and Trust: Local SQLite storage keeps data on the user’s machine, while cryptographic erasure verification and audit trails support compliance controls.
Compliance and Trust: Bayesian trust scores restrict low-trust writes to Core Memory, accelerate low-trust memory decay, and influence retrieval-fusion weights.
Benchmark Results: 100% precision distinguishes full-fidelity from quantized embeddings in the mixed-precision preference benchmark.
Benchmark Results: 68% of recall@10 is preserved when 50% of facts are quantized, while 2-bit embeddings reach 0.801 fidelity at 192× compression.
Benchmark Results: 6.7× discriminative power separates hot from cold memories at day 30, producing hot→4-bit, warm→2-bit, and cold→deleted outcomes.
Benchmark Results: 100% session continuity preserves and recalls all ten tested facts at rank 1 after reopening the engine.
Benchmark Results: 70.4% overall LoCoMo accuracy includes +23.8pp on multi-hop and +12.7pp on adversarial compared with the retrieval baseline.
Benchmark Results: The 4.4pp gap from Paper 2 is concentrated in single-hop and reflects increased seven-channel fusion complexity, which benefits complex queries but dilutes simple-query precision.

11 Limitations and Future Work

The paper identifies boundaries in adaptation, compression, retrieval, platform integration, and model capability, and proposes corresponding directions for future work.

Behavioral learning requires approximately 200 feedback signals before the full adaptive model trains; cold-start retrieval uses fixed channel weights.
Natural-language soft prompts are less powerful than LoRA: they configure preferences but cannot teach new capabilities.
At 2-bit precision, embedding quality degrades significantly, so that tier is appropriate only for archived memories.
Seven-channel fusion introduces noise on simple single-hop queries; query-dependent channel routing is proposed to address the gap.
The zero-friction pipeline is Claude Code-specific, while automatic lifecycle support for other platforms requires per-platform hook integration.
Future directions include hierarchical hyperbolic embeddings, LoRA parameterization, federated private memory, adaptive forgetting calibration, and query-dependent routing.

12 Conclusion

SuperLocalMemory V3.3 completes a local-first cognitive memory system through five contributions spanning quantization, forgetting, retrieval, implicit parameterization, and lifecycle automation. The system reports mathematical validation, production deployment, privacy-preserving operation, and broad availability.

FRQAD achieves 100% precision at distinguishing full-fidelity from quantized embeddings, with zero prior art.
Ebbinghaus Adaptive Forgetting provides 6.7× discriminative power with provable convergence.
7-Channel Cognitive Retrieval achieves 70.4% on LoCoMo in zero-LLM mode, with +23.8pp on multi-hop.
Memory Parameterization implements Long-Term Implicit memory at zero LLM cost.
Zero-Friction Pipeline automates the complete memory lifecycle via a single install.
The system reports a convergent Fokker-Planck system, a valid FRQAD metric, 3,000+ tests, and over 5,000 monthly downloads.It is deployed on npm and PyPI under the Elastic License 2.0.
A single package provides automatic, privacy-preserving cognitive memory across AI coding sessions without cloud services, API keys, or manual commands.