Nexusyn – Long-term memory engine for AI agents (Go, DiskANN, MCP) Nexusyn launched a long-term memory engine for AI agents that scores 81.1% (284/350) on LongMemEval-S and 74.3% (1,476/1,986) on LoCoMo, according to the company's published benchmarks. The Go-based engine combines bi-temporal fact versioning, hybrid retrieval using Reciprocal Rank Fusion over pgvector HNSW and BM25, cross-encoder reranking, and a native Model Context Protocol server at /v1/mcp that connects to Claude Code, Cursor, and Windsurf. Nexusyn says the system returns grounded, source-cited answers and enforces multi-tenant isolation through PostgreSQL row-level security, addressing outdated-fact retrieval and context-window bloat in vector-only agent memory. Stop building agents that forget. Give your LLMs persistent, grounded memory across sessions. Quickstart quickstart-docker-compose • MCP Integration first-class-mcp-integration • Architecture architecture • Benchmarks benchmarks • Cloud vs Self-Hosted self-hosted-vs-nexusyn-cloud • Documentation https://nexusyn.ai/docs Most AI agent frameworks implement memory by taking the last conversation turn, calculating an embedding, and storing it in a vector database. When the agent queries this "memory", it runs a simple nearest-neighbor search. This breaks down in production: 1. Vectors don't understand time: If a user says "I live in Berlin" in January and "I moved to Madrid" in August, both vectors have identical semantic similarity to "Where do I live?" . A pure vector search frequently returns the outdated fact. 2. Context window bloat: Dumping raw matching chunks into the prompt forces the LLM to resolve contradictions, inflating token costs and increasing hallucination rates. 3. No entity understanding: Vector similarity misses relational facts that don't share keywords e.g. connecting a bug report to the architecture decision that caused it . Nexusyn is a dedicated data plane engine designed specifically for agentic long-term memory : - 🕓 Bi-Temporal Versioning: Every fact is stamped with transaction time and valid time valid from / valid to . New statements supersede older ones automatically without deleting history. - 🔀 Hybrid Retrieval RRF : Fuses dense vector similarity pgvector HNSW + BM25 full-text search + entity graph expansion + recency & date-anchor boosts using Reciprocal Rank Fusion. - 🎯 Reranking: High-precision cross-encoder re-scoring of candidate chunks before generation. - 🛡️ Grounded Synthesis: Returns a synthesized, ready-to-use answer that cites verifiable source chunks — so the agent receives answers, not just raw text fragments. - ⚡ MCP-Native: First-class Model Context Protocol server /v1/mcp running over streamable HTTP. Connects directly to Claude Code , Cursor , Windsurf , or custom agent frameworks with zero glue code. - 🏢 Multi-Tenant with Row-Level Security RLS : True isolation at the PostgreSQL layer. A tenant can never see or search another tenant's memories. - ⚙️ Async Pipeline River : Heavy background workloads chunking, batch embedding, entity extraction, and wiki compilation run asynchronously on a Postgres-backed queue. | Capability | Raw Vector DB | Simple Buffer / Window | Nexusyn Engine | |---|---|---|---| | Semantic Vector Search | ✅ | ❌ | ✅ | | BM25 Keyword Search | | ❌ | ✅ | | Bi-Temporal Fact Superseding | ❌ | ❌ | ✅ | | Entity Knowledge Graph GraphRAG | ❌ | ❌ | ✅ | | Grounded Answers with Sources | ❌ Raw chunks only | ❌ | ✅ | | Native MCP Server | ❌ | ❌ | ✅ /v1/mcp | | Database Multi-Tenancy RLS | | ❌ | ✅ Engine-enforced | | Token Cost Efficiency | | ❌ Huge prompts | ✅ Synthesized / Grounded | Nexusyn is evaluated against standard public benchmarks for agent long-term memory under reproducible conditions: | Benchmark | Nexusyn Score | Metric Focus | |---|---|---| | LongMemEval-S | 81.1% 284/350 | Multi-session recall, temporal updates, and preference tracking across long horizons | | LoCoMo | 74.3% 1,476/1,986 | Complex multi-turn conversational reasoning, aggregation, and contradiction resolution | ┌─────────────────────────┐ │ AI Agent / IDE │ │ Claude Code, Cursor, … │ └────────────┬────────────┘ HTTP │ MCP /v1/mcp ▼ ┌────────────────────────────────────────────────────────────────────────┐ │ NEXUSYN ENGINE │ │ │ │ POST /v1/ingest POST /v1/query │ │ │ │ │ │ ▼ ▼ │ │ Deduplication Sub-query Gen │ │ │ │ │ │ Chunker Hybrid Retrieval │ │ │ ├── Vector HNSW │ │ ▼ ├── BM25 Full-Text │ │ River Queue Async ├── Entity Graph │ │ ├── Batch Embeddings └── Date Anchor Boost │ │ ├── Entity & Relation Extraction │ │ │ └── Wiki / Profile Compilation ▼ │ │ │ RRF Fusion │ │ ▼ │ │ │ PostgreSQL 18 pgvector Cross-Encoder Rerank │ │ Row-Level Security Multi-Tenant │ │ │ ▼ │ │ Grounded Answer │ │ Synthesized + Sources │ └────────────────────────────────────────────────────────────────────────┘ Get the complete stack running locally PostgreSQL 17+ with pgvector , the Nexusyn HTTP/MCP API on :8044 , and the River background worker in under 60 seconds: 1. Clone the repository git clone https://github.com/nexusyn/engine.git cd engine 2. Configure environment cp .env.example .env Edit .env with your preferred model provider API keys OpenAI, Anthropic, Gemini, Ollama, Jina, etc. 3. Spin up the containers docker compose up -d 4. Verify health curl http://localhost:8044/health {"status":"ok","time":"..."} Nexusyn runs an HTTP Model Context Protocol MCP server at /v1/mcp . Any MCP-compliant client gets persistent long-term memory tools out of the box. Connect Nexusyn to Claude Code with a single CLI command: claude mcp add nexusyn --transport http \ "http://localhost:8044/v1/mcp" \ --header "Authorization: Bearer