agentdb
github.com/ruvnet/agentdb- Category
- AI Tools
- Rank
- No. 1424Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- ruvnet
- GitHub
- 88 stars
- Date
About
Vector memory that gets smarter every time your agent uses it — feedback-aware retrieval that learns which results actually helped.
What it does
AgentDB packages agent memory into a portable local container holding vectors, indexes, learning state, and an audit trail. It combines approximate vector search with keyword retrieval, episodic records, skills, causal graphs, quantization, and reward-driven ranking. Applications can embed it as a TypeScript library, operate it from a shell, or expose its capabilities to agents through MCP.
Why it's ranked here
The strongest case is breadth without a required server: one package covers storage, retrieval, structured memory, learning, and several integration surfaces. The repository also includes benchmarks, browser builds, tests, and fallback implementations. The verdict stays mixed because the package is alpha, depends heavily on optional native components, and documents broken native attention support.
What's good
The single-container design keeps vectors, indexes, learning state, and audit history portable and usable offline. Hybrid retrieval joins BM25 keyword matching with dense similarity and supports several fusion strategies. The package exposes Node, browser, CLI, and MCP surfaces, while quantization and backend fallbacks give developers explicit choices around memory use, speed, and compatibility.
Tradeoffs
The surface area is unusually large, spanning search, graphs, reinforcement learning, embeddings, synchronization, security, and observability. That raises integration and validation costs. Several accelerators are optional dependencies, the wrapper code says native attention is broken, and installation deliberately ignores post-install failure. License signals also conflict between package metadata and the README badge.
How to use it well
It best fits TypeScript agent builders who want embedded, offline memory with semantic retrieval and explicit feedback signals. Start with the library for application control, the CLI for experiments, or MCP for tool-calling agents. Measure retrieval quality on your own data before enabling compression or learning policies. It does not replace a hosted, shared database service or broader agent orchestration.
Technical notes+
package.json defines an ES module package for Node 18 or newer, exports compiled JavaScript and declarations, installs an agentdb binary, and lists browser, WASM, backend, controller, security, and wrapper entry points. Its postinstall script ends with || true, so native setup failure may be hidden. src/index.ts exposes the broad library API, including controllers, quantization, hybrid search, benchmarking, security, and an LLM router. src/backends/index.ts presents RuVector and HNSWLib implementations behind a shared backend interface. src/browser/index.ts supplies browser feature detection and configuration presets, but its embedded VERSION reports 2.0.0-alpha.2 while package.json reports 3.0.0-alpha.20. src/wrappers/index.ts explicitly routes around alpha packages and labels native attention as broken, using JavaScript fallbacks instead. src/search/index.ts documents BM25 plus vector retrieval with RRF, linear, and max fusion.
Observed
- License
- package.json declares MIT; the README badge presents MIT OR Apache-2.0.
- Primary language
- TypeScript, compiled to ES module JavaScript with declaration files.
- Install surface
- Published as the agentdb npm package; usable through npm installation or npx.
- Interfaces
- TypeScript/JavaScript library, shell CLI, and MCP server with 41 documented tools.
- Runtime support
- Node 18 or newer; README also documents browser WASM, edge runtimes, and offline operation.
- Storage model
- Single .rvf container for vectors, indexes, learning state, and cryptographic audit history.
- Testing surface
- package scripts include Vitest unit tests, browser bundle tests, MCP stdio tests, builds, and bundle verification.
- Backend structure
- Shared vector backend interface with RuVector and HNSWLib implementations plus automatic selection and fallback.
Read from README.md, package.json, src/index.ts, src/simd/index.ts, src/search/index.ts, src/browser/index.ts, src/backends/index.ts, src/wrappers/index.ts, src/embedding/index.ts, src/controllers/index.ts, src/quantization/index.ts, src/compatibility/index.ts, src/observability/index.ts, src/optimizations/index.ts, src/backends/hnswlib/index.ts.
What it can do
Store vector embeddings with feedback tracking
Data vectors and associated metadata → Stored vector memory with feedback mechanisms
Retrieve relevant results based on vector similarity
Query vector or search parameters → Ranked list of similar vectors/documents
Learn from feedback on retrieval results
User feedback on result helpfulness → Updated retrieval rankings and improved future results
Adapt retrieval algorithm based on usage patterns
Historical interaction data and feedback scores → Optimized retrieval model for better accuracy
Track which results were actually helpful to agents
Agent interaction data and outcome metrics → Feedback scores and helpfulness ratings for stored vectors
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.