# CueMap

> CueMap is a Rust memory layer for fast, accurate, and explainable context recall. It turns natural-language memories and queries into inspectable cues, intersects and ranks evidence in memory, and returns the most relevant original memories. v0.7.3 ships bundled local qint8 MiniLM-L3 semantic reranking by default; the edge profile uses a q4 build of the same model.

CueMap is currently documented as version 0.7.3. The Rust engine, its checked-in configuration, API routes, evaluation harnesses, and reports are the source of truth for product claims.

## Start here

- [CueMap home](https://cuemap.dev/): Product overview, live recall trace, architecture, evaluations, ecosystem, and quick start.
- [Quick start](https://cuemap.dev/docs): Install and use CueMap through the CLI, Python or TypeScript SDKs, MCP, Docker, or HTTP.
- MCP exposes repository initialization and status, explicit URL/content/file ingestion, project and memory inspection, project load/unload lifecycle actions, aliases, lexicon administration, and lexical/semantic/hybrid recall.
- [CueMap MCP for Codex](https://cuemap.dev/mcp/codex)
- [CueMap MCP for Claude Code](https://cuemap.dev/mcp/claude-code)
- [CueMap MCP for Google Antigravity](https://cuemap.dev/mcp/antigravity)
- [CueMap MCP for OpenCode](https://cuemap.dev/mcp/opencode)
- [CueMap MCP for VS Code](https://cuemap.dev/mcp/vscode)
- [CueMap MCP for Cursor](https://cuemap.dev/mcp/cursor)
- [HTTP API reference](https://cuemap.dev/docs/api-reference): Current endpoints with sample requests and responses.
- [OpenAPI 3.1 contract](https://cuemap.dev/openapi.yml): Downloadable HTTP API description.
- [GitHub repository](https://github.com/cuemap-dev/cuemap): Source, engine documentation, evaluation code, and reports.

## Product

- [CueMap repository recall trace](https://cuemap.dev/#trace): A source-backed v0.7.3 trace showing real repository ingestion, lexical discovery, local hybrid reranking, and the original returned memories.

## Documentation

- [Automated ingestion](https://cuemap.dev/docs/automated-ingestion): Watch directories, supported formats, structural extraction, and ignore behavior.
- [Server configuration](https://cuemap.dev/docs/configuration): Active TOML settings and configuration precedence.
- [Recall controls](https://cuemap.dev/docs/recall-tuning): Retrieval depth, evidence constraints, expansion, salience, reinforcement, and explanation.
- [Reproduce evaluations](https://cuemap.dev/docs/benchmarks): Commands for LongMemEval, LoCoMo, and BEAM.
- [Storage and access controls](https://cuemap.dev/docs/security): Compression, encryption at rest, and API-key authentication.
- [Project isolation](https://cuemap.dev/docs/projects): Project-scoped writes and explicit cross-project recall.
- [Market heatmap](https://cuemap.dev/docs/market-heatmap): Bounded usage signals and memory ranking.

## Research

- [Performance benchmarks](https://cuemap.dev/research/performance): Wikipedia-based ingestion and recall latency methodology, machine details, controls, and reproduction command.
- [LongMemEval](https://cuemap.dev/research/longmemeval): Raw retrieval, evidence quality, methodology, and full report.
- [LoCoMo](https://cuemap.dev/research/locomo): Retrieval depth, context-token footprint, question types, methodology, and full report.
- [BEAM](https://cuemap.dev/research/beam): Raw retrieval from 128K to 10M, evidence quality, question types, and reproduction commands.

## Use cases

- [Coding agent memory](https://cuemap.dev/use-cases/coding-agent-memory)
- [Long-running assistants](https://cuemap.dev/use-cases/long-running-assistants)
- [Support and incident investigation](https://cuemap.dev/use-cases/support-incident-investigation)
- [Time-sensitive personalization](https://cuemap.dev/use-cases/time-sensitive-personalization)
- [Agent workflow and tool-history memory](https://cuemap.dev/use-cases/agent-tool-history)
- [Decision and organizational memory](https://cuemap.dev/use-cases/decision-organizational-memory)

## Important factual boundaries

- CueMap recall returns stored memories, not an LLM-generated answer.
- Recall explanations expose matched cues and scoring details; they are not chain-of-thought.
- Semantic reranking uses bundled local MiniLM assets; it does not call an external model service.
- Evaluation pages report retrieval metrics over returned memories, not generated-answer accuracy.
- Benchmark diagnostics are labeled separately from normal product-mode recall.