Launching today

CueMap

Launching today

Fast, accurate and inspectable context recall for AI agents

1 follower

Fast, accurate and inspectable context recall for AI agents

1 follower

Visit website

AI Coding Agents

•

Databases and backend frameworks

•

LLM Memory

CueMap is an open-source memory engine for AI agents, built in Rust. It combines deterministic lexical and structural recall with a customized local MiniLM reranker—no external model calls required. At 1M memories, average reads take 2.63ms lexical or 8.36ms hybrid. Raw Hit@20 reaches 96.2% on LongMemEval, 96.1% on LoCoMo, 80.3% on BEAM 1M, and 67.0% on BEAM 10M. Use it through MCP, Python/TypeScript SDKs, Docker, or CLI, with source-backed results and inspectable traces.

Free

Launch tags:Open Source•Artificial Intelligence•GitHub

Launch Team

Zendesk for StartupsUp to 2 years free for startups + AI Agents & Copilot

Promoted

OpenAI

GPT-6 Astra Challenge

What real task does your product handle with GPT-6 Astra?

Maker

CueMap gives GPT-6 Astra persistent, searchable project memory for long-running coding tasks. Through MCP, an Astra-powered agent can index a repository, documentation, requirements, and prior decisions, then retrieve the exact functions, file paths, facts, and history needed for the current task. CueMap performs retrieval locally and returns ranked, source-backed results with inspectable traces. This helps Astra work across large projects and multiple sessions without repeatedly placing the entire repository or conversation history into its context window.

Report

6h ago

Maker

📌

Hello Product Hunt, CueMap is an open-source memory layer for fast, accurate, and explainable context recall. It is written in Rust, runs locally, and makes no external model calls. Candidate discovery stays deterministic. A natural-language query is decomposed into explicit cues, weighted facets, and query-plan labels. CueMap matches these signals against an inverted index to produce a bounded candidate set. In hybrid mode, a bundled local MiniLM model reranks that shortlist. Semantic similarity improves the final ordering without controlling which memories are discovered. Original content and provenance remain attached, while inspectable traces expose how each result was retrieved and scored. On 1M short Wikipedia passages, including HTTP: - Lexical: 2.63ms average / 3.72ms P95 - Hybrid: 8.36ms average / 10.67ms P95 Raw retrieval Hit@20: - LongMemEval: 96.2% - LoCoMo: 96.1% - BEAM 128K: 84.2% - BEAM 1M: 80.3% - BEAM 10M: 67.0% These are retrieval measurements taken before an answer model sees or interprets the returned context. CueMap is available through MCP, Python and TypeScript SDKs, Docker, and CLI. The engine, integrations, evaluation harnesses, benchmark methodology, and reports are open source. It is built for developers who need predictable latency, source-backed results, and direct control over how context is discovered, ranked, expanded, and assembled. Our longer-term goal is to make memory feel like a living system rather than a static search index. As CueMap moves toward v1.0, we’re working toward recall that is fast enough to sit beside inference loops, accurate enough to challenge leading memory systems, and controllable enough to become a multi-request workflow where agents can expand, reorder, and refine evidence before answering. https://cuemap.dev

Report

6h ago

OpenAI

GPT-6 Astra Challenge

What real task does your product handle with GPT-6 Astra?

Maker

Report

6h ago

Maker

📌

Report

6h ago