ICE - Indexless Context Engine

0 reviews

We built ICE at Dopove after watching teams hit the same wall repeatedly: an LLM system works fine in demos, then degrades in production because the model cannot reliably hold, retrieve, and attend to the right context over long workflows. The pain shows up as specific failures, not abstract architecture concerns: - "The agent completed steps 1–5, then forgot the result of step 2." - "We hacked together pgvector, Redis, summaries, and session state — now nobody trusts it." - "If this ever leaks data between customers, we're dead." - "We keep shoving more context in, and quality gets worse." Teams already in production usually respond by stitching together a custom memory stack (pgvector + Redis + session IDs + prompt compression + summaries) and tolerating it. That stack is fragile, expensive to debug, and breaks in ways that are hard to reproduce. ICE is a drop-in infrastructure layer that sits between your existing application and any upstream LLM (OpenAI, Anthropic, Gemini, Ollama). Because it uses the OpenAI-compatible API, there are zero code changes to your application. You keep your SDK, your prompts, your stack. What it handles: - Long-running agent tool-result continuity — agents don't lose earlier outputs in multi-step workflows - Cross-session recall — relevant context from past sessions is retrieved automatically, not re-sent - Kernel-level multi-tenant isolation — PostgreSQL RLS enforced at the DB layer, not the application layer - Sovereign / VPC deployment — for regulated enterprises where data cannot leave the customer's boundary - No framework rewrite needed — works with or without LangGraph, LlamaIndex, etc. Benchmarks (v2.6.755, 32GB RAM, production infra): - Semantic recall latency: 15ms - 687 req/sec throughput - 10,000 concurrent sessions per node We are a member of NVIDIA Inception Program. If your team is already in production with long-lived agents, persistent enterprise copilots, or multi-tenant AI products - and the memory layer is where things break - I'd genuinely like to hear how you're handling it today. Happy to share docs or set up a 20-minute call.