DSM: Dynamic Segmented Memory

"Infinite memory for all LLMs." Developed by Nare Labs

DSM (Dynamic Segmented Memory) is a high-performance memory engine that enables models to reason over datasets with millions of tokens. It replaces dense attention bottlenecks with a hierarchical, graph-based associative memory architecture.

---

The Architecture: Triplet State (S, T, G)

DSM organizes knowledge into three interconnected layers:

• S (Segments): Atomic units of text with semantic embeddings.
• T (Hierarchy): A dynamic category tree used for high-level beam-search routing.
• G (Graph): A semantic graph preserving associative links between related segments.

Why DSM vs. Standard RAG?

Standard RAG (Retrieval-Augmented Generation) often treats context as a flat list of chunks, leading to several limitations that DSM solves:

Challenge	Standard RAG	DSM Engine
Context Fog	Chunks are retrieved in isolation.	Graph edges preserve logical flow and dependencies.
Search Speed	Linear vector search over $N$ chunks.	Hierarchical routing ($T$) enables $O(\log N)$ pruning.
Associativity	Cannot "hop" to related concepts.	Graph expansion ($G$) finds connected "needles" automatically.
Organization	Flat database.	Dynamic Tree ($T$) mirrors the semantic structure.

---

Performance & Efficiency

Tested on consumer-grade hardware to demonstrate extreme efficiency for Small Language Models (SLMs):

Metric	Dense Attention (1M tokens)	DSM Engine
Compute Ops	~26,000,000,000	65,536
Retrieval Latency	Minutes	~430ms
Efficiency Boost	1x	401,075x faster
Needle Accuracy	~60-80%	100%

Benchmarks performed using Qwen-2.5-1.5B to prove that DSM can grant "large-model" reasoning capabilities to even the smallest architectures.

---

🛠️ Retrieval Flow

graph TD
    UserQuery[User Query] --> Router{DSM Router}
    Router -->|1. Beam Search| CatTree[Category Tree]
    Router -->|2. k-NN| VectorIndex[Semantic Index]
    Router -->|3. Hops| MemoryGraph[Associative Graph]
    
    CatTree --> Candidates
    VectorIndex --> Candidates
    MemoryGraph --> Candidates
    
    Candidates --> Scorer[Priority Scorer]
    Scorer -->|Ensemble Scoring| ActiveContext[Active Context]
    ActiveContext -->|Injection| LLM((Any LLM))

Loading

💻 Quick Start

Installation

git clone https://github.com/narelabs/dsm
cd dsm
pip install -e .

Python API

from dsm import DynamicSegmentedMemory

# Initialize Engine
memory = DynamicSegmentedMemory(".dsm/storage.json")

# Ingest Content
memory.write(
    "The core logic resides in dsm/memory.py. It handles hierarchical routing.",
    category_path="Software / Architecture / Core",
    importance=0.9
)

# Retrieve Active Context
ctx = memory.active_context("How does DSM route queries?", k=3)

print(f"Context size: {ctx.estimated_tokens} tokens")
print(ctx.context_text)

📜 License

Released under the MIT License. Created by Nare Labs. narelabs.com

---

Developed for the next generation of autonomous AI engineers.