# HotKV > HotKV is a multi-threaded, Redis-compatible in-memory data store written in C++20, with LLM caching, agent memory, RAG, and vector search built into the engine. It speaks RESP2 and RESP3, implements 300 core commands plus more than 170 AI and enterprise commands, and adds key tagging, nanosecond TTLs, and optional value compression. HotKV is commercial software, available as HotKV Cloud (managed Starter, Performance, and Scale plans) or with an on-premise license. Key facts: - Multi-threaded I/O and command execution against a sharded store. In an October 2026 benchmark on a 160-vCPU Azure machine (256-byte values, pipeline depth 30), HotKV, built from source, had higher throughput than Redis 8.10.2 and Valkey 9.1.2 on all 12 commands tested, and higher than Dragonfly 2.0.2 on 11 of 12 (Dragonfly was level on PING). - Works with existing Redis client libraries and tools over RESP2/RESP3. Runs on Linux, x86_64 and ARM64. - Managed plans: Starter (1 vCPU, 4 GB RAM, 10 GB storage, 100 GB traffic/month), Performance (2-96 vCPUs, 16-768 GB RAM, 250 GB-8 TB traffic/month, all AI and LLM modules), Scale (4-384 vCPUs, 31 GB-3 TB RAM, 512 GB-25 TB traffic/month, plus encryption at rest, audit log, and multi-tenancy). Traffic top-up packs on every plan; monthly or yearly billing; plans upgrade in place. - AI command families (Performance and Scale plans): prompt cache, semantic cache, negative cache, tool-call cache, rerank-result cache, agent memory, RAG pipelines, embedding de-duplication, LLM budgets and cost tracking, time series, feature store, and rate limiting. - In a benchmark with real Anthropic API calls on Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 (3,000 requests each), HotKV cut input plus output tokens by about 70% on every model and answered at least 80% of requests from cache once warm. A judge model found that at the default 0.81 similarity threshold a cached answer fit the exact question 6 to 11 percentage points less often than a fresh answer, and that on Sonnet 5.5, from a threshold of 0.88 upward, it did not, at 61% to 64% lower cost. - Official client libraries for Python, Node.js, Go, Rust, Java, C#, C, C++, PHP, Ruby, Perl, and Elixir, open source under the Apache License 2.0. - Open-source AI tools under the Apache License 2.0: an MCP server for AI assistants (hotkv-mcp), a LangGraph checkpointer and memory store (langgraph-checkpoint-hotkv), and a caching gateway for OpenAI-compatible APIs and Anthropic's Messages API (hotkv-gateway). ## Product - [Features](https://hotkv.com/features): Full feature breakdown, including the comparison against Redis, Valkey, and Dragonfly. - [Performance](https://hotkv.com/performance): Benchmark methodology and results. - [Architecture](https://hotkv.com/architecture): The multi-threaded execution model, sharding, and persistence design. - [AI-driven workloads](https://hotkv.com/ai): RAG, semantic caching, agent memory, and the rest of the AI command set. - [Use cases](https://hotkv.com/use-cases): 31 real-world use cases, each listed below with a step-by-step walkthrough. ## Use cases Each page walks through one workload command by command, with real HotKV commands and replies in the documented format. The Markdown copies hold the same content. - [Build an AI Customer Support Assistant on HotKV](https://hotkv.com/use-cases/ai-support-assistant.md): A production AI support assistant on one engine, with prompt and semantic caching, RAG over your help center, conversation memory and per-tenant budgets. - [Caching Gateway for OpenAI and Anthropic APIs | HotKV](https://hotkv.com/use-cases/llm-gateway.md): hotkv-gateway is an open-source caching proxy for LLM APIs, with exact caching, semantic caching for OpenAI-compatible chat, stream replay and coalescing. - [Semantic Cache for LLM Apps with HotKV](https://hotkv.com/use-cases/semantic-caching.md): Cut LLM cost and latency with HotKV's SCACHE commands. Cache answers by embedding similarity so rephrased prompts hit the cache, with per-namespace stats. - [LLM Prompt Cache with Token Savings | HotKV](https://hotkv.com/use-cases/prompt-caching.md): HotKV's PCACHE commands cache LLM responses for exact prompts, record model token usage, and report hits, lookup latency and money saved per namespace. - [Prompt Registry with Versioning and Rollback | HotKV](https://hotkv.com/use-cases/prompt-management.md): Store LLM prompt templates in HotKV with PROMPT.PUT, read the active version at runtime, and roll back instantly with PROMPT.SETACTIVE. No redeploy needed. - [RAG Pipelines with Hybrid Vector and BM25 Search](https://hotkv.com/use-cases/rag-knowledge-assistants.md): Build retrieval-augmented generation on HotKV. RAG.INGEST chunks documents and RAG.SEARCH filters by metadata and runs hybrid vector and BM25 search. - [Vector Similarity Search with HotKV](https://hotkv.com/use-cases/vector-search.md): Run vector similarity search on HotKV with VSIM commands, HNSW indexes, cosine, L2 or inner product, and tag filters right next to your product data. - [Real-Time Recommendation Engine on HotKV](https://hotkv.com/use-cases/recommendations.md): Serve real-time recommendations from HotKV with feature store reads, vector similarity, Top-K trending and Bloom filters to skip items a user has seen. - [AI Agent Memory for LLM Apps with HotKV](https://hotkv.com/use-cases/ai-agent-memory.md): Give AI agents persistent memory on HotKV with the AMEM commands, covering conversation windows, semantic recall, working memory and an episodic log. - [Multi-Agent Orchestration on HotKV](https://hotkv.com/use-cases/multi-agent-orchestration.md): Coordinate AI agents with HotKV using stream consumer groups for task handoff, shared working memory, a shared tool cache and an episodic action log. - [Cache AI Agent Tool Calls with HotKV](https://hotkv.com/use-cases/agent-tool-caching.md): HotKV's tool cache memoizes deterministic tool and function calls made by AI agents, with TTLs, entry limits and statistics on requests and cost saved. - [LangGraph Checkpointer and Store for HotKV](https://hotkv.com/use-cases/langgraph-checkpoints.md): langgraph-checkpoint-hotkv gives LangGraph agents a HotKV checkpointer and long-term store. It passes LangGraph's conformance suite, no license needed. - [HotKV MCP Server, Memory for Claude and Cursor](https://hotkv.com/use-cases/mcp-memory.md): hotkv-mcp is an open-source MCP server that gives AI assistants conversation memory, notes, event logs, prompt templates and document search in HotKV. - [LLM Token Budgets and Cost Tracking with HotKV](https://hotkv.com/use-cases/llm-cost-governance.md): Enforce LLM token and spend budgets atomically with HotKV's LLMBUDGET commands, track usage per tenant with LLMSTATS and deduplicate embeddings by hash. - [Cache Validated LLM Output and Moderation Results | HotKV](https://hotkv.com/use-cases/ai-guardrails-caching.md): HotKV's guard cache stores LLM responses that passed schema or guardrail checks, and its negative cache remembers refusals, so repeats skip the model. - [Real-Time ML Feature Store on HotKV](https://hotkv.com/use-cases/ml-feature-store.md): HotKV's feature store keeps typed feature groups for low-latency inference, written with FEATURE.SET and read with FEATURE.GET and FEATURE.BATCH. - [Redis-Compatible In-Memory Caching with HotKV](https://hotkv.com/use-cases/caching.md): Use HotKV as a Redis-compatible cache with cache-aside reads, TTLs down to nanoseconds, optional compression, tag-based invalidation and multi-threading. - [Web and API Session Store with HotKV](https://hotkv.com/use-cases/session-management.md): Run web and API sessions on HotKV with automatic expiry and key tagging to find or revoke sessions by user, role or device, with no secondary index. - [API Rate Limiting with GCRA and Sliding Windows](https://hotkv.com/use-cases/rate-limiting.md): Protect your API with HotKV. GCRA.LIMIT gives smooth limits with bursts, and named limiters offer sliding window, token or leaky bucket, all atomic. - [Real-Time Leaderboards with HotKV](https://hotkv.com/use-cases/leaderboards.md): Build leaderboards on HotKV sorted sets, with ZADD and ZINCRBY for scores, ZREVRANGE for top lists and ZREVRANK for a player's position. - [Background Job Queues on Lists with HotKV](https://hotkv.com/use-cases/job-queues.md): Run background job queues on HotKV with LPUSH, BRPOP and BLMOVE. Hand each job to a processing list so none is lost if a worker crashes. - [Event Streaming with HotKV Streams](https://hotkv.com/use-cases/event-streaming.md): Use HotKV streams for event-driven systems, with XADD to publish, XREADGROUP to share work, XACK to acknowledge and XAUTOCLAIM to recover messages. - [Real-Time Chat and Notifications with HotKV Pub/Sub](https://hotkv.com/use-cases/pub-sub-messaging.md): Build real-time chat, presence and notifications on HotKV with PUBLISH and SUBSCRIBE, pattern subscriptions, sharded pub/sub and streams for history. - [Distributed Locks with SET NX and DELEX in HotKV](https://hotkv.com/use-cases/distributed-locks.md): Implement distributed locks on HotKV with SET NX PX and release them safely with DELEX IFEQ, so a worker can only remove its own lock. No Lua needed. - [Flash Sale Inventory and Checkout with HotKV](https://hotkv.com/use-cases/flash-sale-inventory.md): Handle flash sales on HotKV with atomic stock counters, cart holds that expire, bot limits with GCRA and an order stream for fulfilment. - [Real-Time Fraud Detection Signals with HotKV](https://hotkv.com/use-cases/fraud-detection.md): Compute fraud signals in real time on HotKV with velocity counters, Bloom filters of known-bad cards, Count-min sketches and feature store reads. - [Geospatial Search and Nearby Matching with HotKV](https://hotkv.com/use-cases/geospatial.md): Track live locations and find what is nearby with HotKV, using GEOADD and GEOSEARCH by radius or box, for ride-hailing, delivery and store locators. - [IoT and Metrics Time Series with HotKV](https://hotkv.com/use-cases/iot-time-series.md): Store IoT and metrics time series in HotKV with TS.CREATE retention and labels, TS.ADD samples, TS.DOWNSAMPLE rules and TS.AGG bucket aggregation. - [Real-Time Analytics, Counters and Trending with HotKV](https://hotkv.com/use-cases/real-time-analytics.md): Count and rank events in real time with HotKV, using INCR counters, sorted sets for trending lists, HOTKEYS hot key detection and Prometheus metrics. - [HyperLogLog, Bloom, Top-K and Count-Min Sketch in HotKV](https://hotkv.com/use-cases/probabilistic-counting.md): Count unique users with HyperLogLog, deduplicate with Bloom filters, track top items with Top-K and estimate frequencies with Count-min sketches in HotKV. - [Multi-Tenant SaaS on HotKV with ACL Isolation](https://hotkv.com/use-cases/multi-tenant-saas.md): Run many tenants on one HotKV cluster. ACL key patterns confine each tenant to its own keys, and a tenant registry records users and quota values. ## Docs - [Documentation](https://docs.hotkv.com): Quick starts, the command and configuration reference, data types, clustering, persistence, security and the client libraries. It has its own llms.txt at https://docs.hotkv.com/llms.txt. ## Pricing - [Pricing](https://hotkv.com/pricing): Starter, Performance, and Scale managed plans with bundled traffic, and on-premise licensing. - [Cloud console](https://console.hotkv.com): Sign up, launch, and manage instances. ## Blog - [Blog](https://hotkv.com/blog): Announcements and engineering posts. - [Norbert E. Boros](https://hotkv.com/author/norbert): Founder and CEO of HotKV Ltd, who wrote HotKV and writes the blog. - [Redis benchmark: HotKV vs Valkey and Dragonfly in a three-node cluster](https://hotkv.com/blog/redis-benchmark-cluster.md): A Redis benchmark of HotKV, Redis, Valkey and Dragonfly as a three-node cluster on AWS: 9 commands, pipeline depth 100, and where HotKV was behind. - [AI agent memory: four types and how to store them](https://hotkv.com/blog/ai-agent-memory.md): AI agent memory is more than a long context window. The four kinds an agent needs, how to store each and what to keep out, with HotKV commands. - [MCP server for agent memory: connect Claude Code to HotKV](https://hotkv.com/blog/mcp-server-for-agent-memory.md): An MCP server gives Claude Code or Cursor memory in HotKV through 28 fixed tools. Install it, connect it and limit it with read-only mode and namespaces. - [Prompt caching vs semantic caching: what each one saves](https://hotkv.com/blog/prompt-caching-vs-semantic-caching.md): Prompt caching makes a repeated prefix cheaper. Semantic caching removes the model call. How each works, what we measured and which to use when. - [Vector search: HotKV vs Redis 8 vector sets, from 1 to 100 clients](https://hotkv.com/blog/vector-search-benchmark.md): HotKV and Redis 8.10.2 vector search on 100,000 and 999,000 embeddings from 1 to 100 clients: throughput, latency, recall, loading speed, memory and limits. - [HotKV cut Claude token use by 70% in our Banking77 benchmark](https://hotkv.com/blog/token-savings-benchmark.md): Four Claude models, 3,000 support requests each: HotKV caching cut model calls, tokens and API cost by about 70%. Results and answer-quality trade-offs. - [Introducing HotKV](https://hotkv.com/blog/introducing-hotkv.md): Why I built HotKV, a multithreaded, Redis-compatible data store written from scratch in modern C++, what it offers today and how it did in our benchmark. ## Legal - [Terms and Conditions](https://hotkv.com/legal/terms-and-conditions) - [Privacy Policy](https://hotkv.com/legal/privacy-policy) ## Optional - [Everything in one file](https://hotkv.com/llms-full.txt): every blog post and use case page above, as Markdown.