I'm Saran S, founder of Dopove.  Always open to connecting with developers, enterprises, and fellow builders.

projects

devto

medium

substack

productHunt

Saran S • Peerlist

GLACIER: The Open-Source Memory Engine for Mamba and Next-Gen AI AgentsGLACIER is a revolutionary open-source project by Dopove Private Limited, architected by Saran S, that finally solves the fundamental memory limitations of State Space Models (SSMs) like Mamba, and addresses the scaling challenges of traditional Transformer-based RAG systems. It is designed to empower developers to build truly intelligent, persistent, and efficient AI applications capable of engaging in long, coherent conversations and complex agentic workflows.The Problem GLACIER Solves: 1. Mamba's Amnesia &amp; Context Rot: While Mamba models offer unparalleled O(1) inference speed, they suffer from "context rot"—a lack of persistent memory. Their hidden state continuously overwrites itself, causing them to "forget" early conversation details in long sessions. GLACIER gives Mamba an external hippocampus, ensuring crucial facts are never lost. 2. Transformer Scaling Limitations: Traditional Transformer models, when combined with RAG, must store and process the entire conversation history in their KV-cache. This leads to linear token growth and quadratic (O(N^2)) latency scaling, making long conversations prohibitively expensive and slow.How GLACIER Works: A Two-Pillar Architecture GLACIER introduces two core innovations: * ICE-Lite (Infinite Context Engine): This lightweight virtual memory management layer acts as an "external hippocampus" for Mamba. It transparently intercepts model prompts, performs rapid semantic retrieval from a persistent episodic ledger, and "pages in" only the most relevant context into Mamba's active window. This decouples long-term memory from the model's transient state. * Temporal-RAG: A novel, time-aware reranking layer built directly into ICE-Lite. Beyond mere semantic similarity, Temporal-RAG classifies documents by validity (VALID, TEMPORAL, EXPIRED) and kind (STATIC, VERSIONED, EVENT). It intelligently applies time-decay scoring, recency weighting, and semantic relevance thresholds to ensure Mamba always receives the freshest, most pertinent information, actively demoting stale or outdated facts.Key Features &amp; Benefits for Developers: * Persistent &amp; Infinite Memory: Mamba agents can now remember facts, instructions, and entire conversation histories across hundreds of turns and even application restarts. No more "Agentic Amnesia." * Massive Token Efficiency: GLACIER drastically reduces the token footprint. Our empirical benchmarks show GLACIER is 16.7x more memory efficient than a Transformer baseline, maintaining a constant ~300 tokens in the active prompt compared to ~4900 tokens for a Transformer at Turn 100. * Constant O(1) Inference Latency: By keeping Mamba's active context small and relevant, GLACIER preserves Mamba's core speed advantage, delivering predictable, flat inference latency regardless of conversation length. * Advanced Agentic Capabilities: Native support for tool-calling (with "JSON pinning" to remember tool outputs), multi-step "Turbo-Stitching" for seamless long-form generation, and one-line ingestion of files and codebases. * Guaranteed Reproducibility with Docker: Includes a Dockerfile for easy setup, ensuring a perfectly configured CUDA environment, eliminating complex Mamba compilation issues on diverse machines. * Empirical Benchmarks: Backed by real-world data and visualizations proving its performance against both vanilla Mamba and Transformer baselines.GLACIER empowers you to: * Build Mamba-powered agents that are truly coherent, consistent, and context-aware over extended interactions. * Drastically reduce operational costs by minimizing token usage and maintaining predictable latency. * Innovate with next-generation AI applications that were previously impossible due to memory limitations.Join the movement! GLACIER is open-source and ready for the community to build the future of memory-augmented SSMs. Explore the code, run the benchmarks, and contribute to pushing the boundaries of AI capabilities.

Glacier By Dopove

<h3>🧊 ICE by Dopove: The Virtual Memory Manager for LLMs</h3>Building AI agents is easy. Keeping them coherent in production is a nightmare.If you are building long-lived agents or enterprise copilots, you’ve likely hit the same wall: the LLM works perfectly in the demo, but degrades in production because it can’t reliably hold, retrieve, and attend to the right context over a long workflow.Right now, most teams respond by duct-taping together a custom memory stack—Postgres + pgvector + Redis + session IDs + summarization loops. It’s fragile, expensive to debug, and leads to specific, bleeding-neck failures:<ul><li>Agents hit Step 10 and forget the tool output from Step 2.</li><li>You shove more context into the prompt, your token bills explode, and the model still suffers from "Lost in the Middle" amnesia.</li><li>You live in constant fear of multi-tenant data leakage.</li></ul>Try ICE (Indexless Context Engine).ICE is a drop-in Virtual Memory Management (VMM) layer that sits between your existing application and any upstream LLM (OpenAI, Anthropic, Gemini, Ollama).The best part? Zero migration friction. ICE is 100% API compatible. You literally just change your base URL. You keep your SDKs, your prompts, and your orchestration frameworks (LangGraph, LlamaIndex, etc.).What ICE handles under the hood:<ul><li>Cures Agentic Amnesia: ICE natively understands tool calls. It pins critical, recent tool-results to the active window and safely archives older steps to a Semantic Ledger. Multi-step agents never lose their train of thought.</li><li>BYODB (Bring Your Own Database): Don't migrate your data. Point ICE at your existing PostgreSQL (pgvector) and Redis clusters. You own the storage and the infrastructure; we just provide the memory OS.</li><li>Glass-Box Observability: Engineers hate black boxes. ICE natively exports OpenTelemetry (OTel) traces and includes a dry-run API. You can see the exact token receipts and injected context before it ever hits the LLM. You never lose control of the prompt.</li><li>Precision Paging: We don't do dumb summarization. ICE dynamically pages raw, high-signal context into the prompt exactly when the agent needs it, keeping your physical token count ruthlessly low while guaranteeing perfect historical recall.</li><li>Kernel-Level Multi-Tenancy: Isolation isn't an app-level afterthought. ICE enforces multi-tenancy at the database layer via PostgreSQL Row-Level Security (RLS). Cross-tenant leakage is mathematically impossible.</li><li>Sovereign / VPC Deployment: Built for regulated enterprises where your data and memory state cannot leave your own boundary.</li></ul>Planned Future Features:We are actively building the next generation of LLM physics to completely eliminate token bottlenecks and compute waste.<ul><li>True Infinite Input (Zero-Token Cost): We are moving beyond text prompts. ICE is developing direct KV-Cache memory mapping. Soon, models won't "read" context; ICE will inject it directly into the model's neural state, unlocking instant 100-Billion token horizons with zero input API costs.</li><li>Multi-Tenant Batch Routing: To slash enterprise inference bills, ICE will aggregate queries across your entire application and process them in single, massive inference batches—while using our strict VMM isolation to guarantee tenant data never crosses streams.</li><li>Hardware Defiance: Our ultimate goal is O(1) memory mapping efficiency, allowing you to run 1,000 concurrent, memory-dense AI sessions on a single edge device without degrading performance.</li></ul>If you are shipping persistent enterprise copilots or multi-tenant AI products, stop rebuilding brittle Postgres/Redis memory workarounds.Read the docs, contact us, and drop ICE into your stack today: 👉 <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="https://dopove.com/ice">https://dopove.com/ice</a>