[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:metronix-memory":3},"\u003Ch1>Metronix Memory\u003C\u002Fh1>\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fmtrnix\u002Fmetronix-memory\u002FHEAD\u002Fdocs\u002Fmetronix-banner.svg\" alt=\"Metronix Memory\" width=\"600\" \u002F>\n\u003C\u002Fp>\u003Cp>\u003Cstrong>Self-hosted memory infra for AI agents — MCP-native, local-model friendly: hybrid RAG + temporal knowledge graph & ontology layer, durable memory, freshness checks, agent-scoped context.\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Metronix gives agents a memory backend they can actually call: ingest files and SaaS knowledge, retrieve with dense + sparse + graph context, store durable facts and preferences per agent, and keep long-lived knowledge fresh as projects change.\u003C\u002Fp>\n\u003Ch2>What You Get\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>Durable memory for every agent\u003C\u002Fstrong> — store facts, preferences, and pinned context with\nworkspace and agent scoping.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A knowledge backend your agents can use\u003C\u002Fstrong> — ingest files and connected SaaS sources, then\nretrieve hybrid dense, sparse, and graph context with source citations.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Control over your data and models\u003C\u002Fstrong> — run the full stack yourself with Docker, bundled\nlocal models, and optional external answer generation.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>One integration point for your tools\u003C\u002Fstrong> — connect MCP-native agents now, or use the REST API\nand native Hermes memory provider where they fit best.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>Quick Start\u003C\u002Fh2>\n\u003Cp>Start the local Docker stack, confirm that it is live, then connect your agent:\u003C\u002Fp>\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fmtrnix\u002Fmetronix-memory\u002FHEAD\u002Fdocs\u002Fmetronix-agent-memory-demo.gif\" alt=\"Metronix demo: an agent remembering across sessions\" width=\"720\" \u002F>\n\u003C\u002Fp>\u003Cpre>\u003Ccode class=\"language-bash\">git clone https:\u002F\u002Fgithub.com\u002Fmtrnix\u002Fmetronix-memory.git\ncd metronix-memory\ncp .env.example .env\nprintf '\\nMETRONIX_MCP_API_KEY=%s\\n' \"$(openssl rand -hex 32)\" >> .env\ndocker compose up -d --build\ncurl http:\u002F\u002Flocalhost:8000\u002Fhealth\n# {\"status\":\"ok\"}\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>Then follow \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmtrnix\u002Fmetronix-memory\u002Fblob\u002FHEAD\u002Fconnecting_to_agent.md\" rel=\"nofollow ugc noopener\">Connecting To An Agent\u003C\u002Fa> to give your MCP client durable\nmemory. For the release installer, configuration, and troubleshooting, continue to\n\u003Ca href=\"#install\" rel=\"nofollow ugc noopener\">Install\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>\u003Ca href=\"#install\" rel=\"nofollow ugc noopener\">Install\u003C\u002Fa>\u003C\u002Fstrong> | \u003Ca href=\"#choose-your-runtime-guide\" rel=\"nofollow ugc noopener\">Runtime Guides\u003C\u002Fa> | \u003Ca href=\"#benchmarks\" rel=\"nofollow ugc noopener\">Benchmarks\u003C\u002Fa> | \u003Ca href=\"#documentation\" rel=\"nofollow ugc noopener\">Docs\u003C\u002Fa>\u003C\u002Fp>\n\u003Chr \u002F>\n\u003Ch2>Why Not Just...\u003C\u002Fh2>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Option\u003C\u002Fth>\n\u003Cth>What it gives you\u003C\u002Fth>\n\u003Cth>What Metronix adds\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Vector DB\u003C\u002Ftd>\n\u003Ctd>Similarity search over embedded chunks\u003C\u002Ftd>\n\u003Ctd>Ingestion, MCP tools, durable agent memory, sparse retrieval, graph context, and operational APIs\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Long context\u003C\u002Ftd>\n\u003Ctd>More tokens in one prompt\u003C\u002Ftd>\n\u003Ctd>Persistent memory across sessions, agent\u002Fworkspace scoping, retrieval, and freshness checks\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Chat history\u003C\u002Ftd>\n\u003Ctd>Transcript recall\u003C\u002Ftd>\n\u003Ctd>Structured facts, preferences, pinned memory, temporal knowledge, and reusable context for any MCP-native agent\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Ch2>Benchmarks\u003C\u002Fh2>\n\u003Cp>Directional N=1 results under \u003Ccode>benchmark-protocol v1.0\u003C\u002Fcode>: same answer model (\u003Ccode>deepseek-v4-flash\u003C\u002Fcode>, T=0), same blind judge (\u003Ccode>deepseek-v4-pro\u003C\u002Fcode>), same volume, and both retrieval + end-to-end layers.\u003C\u002Fp>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Benchmark\u003C\u002Fth>\n\u003Cth>Scope\u003C\u002Fth>\n\u003Cth>Layer B result\u003C\u002Fth>\n\u003Cth>Retrieval \u002F signal\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>LoCoMo\u003C\u002Ftd>\n\u003Ctd>1,982 \u002F 1,982 QA pairs\u003C\u002Ftd>\n\u003Ctd>\u003Cstrong>52.8%\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Recall@10 \u003Cstrong>85.3%\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>LongMemEval-S\u003C\u002Ftd>\n\u003Ctd>500 \u002F 500 questions\u003C\u002Ftd>\n\u003Ctd>\u003Cstrong>59.0%\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Recall@10 \u003Cstrong>95.4%\u003C\u002Fstrong>; reproducible harness in \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmtrnix\u002Fmetronix-memory\u002Fblob\u002FHEAD\u002Fbenchmarks\u002Flongmemeval\" rel=\"nofollow ugc noopener\">benchmarks\u002Flongmemeval\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>MemoryAgentBench\u003C\u002Ftd>\n\u003Ctd>2,800 \u002F 2,800 tasks\u003C\u002Ftd>\n\u003Ctd>\u003Cstrong>63.6%\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Accurate Retrieval \u003Cstrong>84.7%\u003C\u002Fstrong>; EventQA blended \u003Cstrong>86.8%\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>EventQA\u003C\u002Ftd>\n\u003Ctd>MAB EventQA 65K + 131K\u003C\u002Ftd>\n\u003Ctd>\u003Cstrong>86.8%\u003C\u002Fstrong> blended\u003C\u002Ftd>\n\u003Ctd>98.0% at 65K; 94.8% at 131K\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>BEAM 100K\u003C\u002Ftd>\n\u003Ctd>400 \u002F 400 questions\u003C\u002Ftd>\n\u003Ctd>\u003Cstrong>32.1%\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Recall@10 2.9%; Layer B is the meaningful figure for this tier\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Cp>Metronix leads the equal-conditions comparison on LoCoMo and MemoryAgentBench, while Mem0 leads narrowly on LongMemEval-S and BEAM 100K. The recurring pattern is retrieval ahead of generation: relevant evidence is usually found, but answer synthesis, conflict resolution, and preference following remain the hard parts.\u003C\u002Fp>\n\u003Ch2>Integrations\u003C\u002Fh2>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Agent\u002Fruntime\u003C\u002Fth>\n\u003Cth>Path\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Hermes\u003C\u002Ftd>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmtrnix\u002Fhermes-memory-metronix\" rel=\"nofollow ugc noopener\">Native memory provider\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmtrnix\u002Fmetronix-memory\u002Fblob\u002FHEAD\u002Fdocs\u002Fintegrations\u002Fhermes-agent.md\" rel=\"nofollow ugc noopener\">MCP guide\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Cursor\u003C\u002Ftd>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmtrnix\u002Fmetronix-memory\u002Fblob\u002FHEAD\u002Fdocs\u002Fintegrations\u002Fcursor.md\" rel=\"nofollow ugc noopener\">Cursor guide\u003C\u002Fa>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Claude Desktop\u003C\u002Ftd>\n\u003Ctd>[Claude Desktop guide](docs\u002Fintegrations\u002Fclaude-des\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n",1788219548897]