My Hermes Agent setup runs on a headless Ubuntu server with no usable local GPU. The existing long-term memory backend, QMD, is a capable semantic-search engine, but it embeds markdown with GGUF models through I wanted something that could:<\/p>\n I evaluated three systems:<\/p>\n I used a mix of representative notes: a system configuration file, a user-profile file, a contact list, and several long-form markdown creative documents. Names, addresses, and phone numbers in the contact sample were replaced with placeholders before any reporting.<\/p>\n For each candidate I measured:<\/p>\n QMD and MemPalace both do good work, but they are embedding-first. On a CPU-only box the embedding step dominates every operation. Mnemosyne’s default mode avoids that entirely by using SQLite FTS5. Search is genuinely sub-second and RAM stays under 100 MB.<\/p>\n The trade-off is recall quality. In pure keyword mode, Mnemosyne can rank a profile note above a note that is literally about<\/em> the queried topic, because keyword frequency wins over semantic intent. Enabling remote embeddings fixes that: queries like “QMD memory” returned the correct document first, and music-related queries surfaced the right creative files.<\/p>\n The old QMD collection had grown to about 244 files and 4.1 MB, dominated by multiple drafts of two novels and their compiled full manuscripts. After reviewing an inventory I removed the duplicate and draft material, keeping only a small set of creative notes. The remaining files were migrated into Mnemosyne as global-scope memories with Ollama-generated embeddings.<\/p>\n I then:<\/p>\n The gateway now starts with these environment variables baked into its systemd unit:<\/p>\n This keeps all memory data local in SQLite while sending only embedding requests to the GPU box. A live test confirmed the provider is active and recall works: a newly stored memory was retrieved correctly by semantic query.<\/p>\n For a CPU-only Hermes deployment that still wants semantic memory, Mnemosyne with a remote Ollama embedding endpoint is the sweet spot. It preserves local speed and privacy, keeps RAM tiny, and only uses remote compute for the parts that actually need a GPU. QMD and MemPalace are better suited to machines that can run embeddings locally without pain.<\/p>\n","protected":false},"excerpt":{"rendered":" Why I looked for a new memory backend My Hermes Agent setup runs on a headless Ubuntu server with no usable local GPU. The existing long-term memory backend, QMD, is a capable semantic-search engine, but it embeds markdown with GGUF models through node-llama-cpp. On a CPU that becomes the bottleneck: every re-indexing batch and every<\/p>\nnode-llama-cpp<\/code>. On a CPU that becomes the bottleneck: every re-indexing batch and every hybrid query carries the overhead of loading and running local embedding and reranking models.<\/p>\n\n
The candidates<\/h2>\n
\n
How I tested<\/h2>\n
\n
Key results<\/h2>\n
\n\n
\n \nSystem<\/th>\n Local RAM<\/th>\n Store latency<\/th>\n Recall latency<\/th>\n Notes<\/th>\n<\/tr>\n<\/thead>\n \n QMD (local GGUF)<\/td>\n Moderate\u2013high<\/td>\n Slow on CPU<\/td>\n Slow on CPU<\/td>\n Best hybrid quality, but CPU-bound<\/td>\n<\/tr>\n \n MemPalace<\/td>\n ~800 MB\u20131.5 GB<\/td>\n Embedding-bound<\/td>\n Fast after embedding<\/td>\n Agent-retention focused, heavier for a raw markdown archive<\/td>\n<\/tr>\n \n Mnemosyne (no embeddings)<\/td>\n ~84 MB<\/td>\n ~0.2\u20130.4 s<\/td>\n ~0.03\u20130.06 s<\/td>\n Keyword-only, very fast, lower semantic recall<\/td>\n<\/tr>\n \n Mnemosyne + remote Ollama<\/td>\n ~87 MB<\/td>\n ~0.5\u20132.8 s<\/td>\n ~0.1\u20131.3 s<\/td>\n Semantic search offloaded to GPU; local footprint stays tiny<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n What the numbers mean<\/h2>\n
Migration and cleanup<\/h2>\n
\n
mnemosyne-hermes<\/code> into the active Hermes virtual environment,<\/li>\n~\/.hermes\/plugins\/mnemosyne<\/code>,<\/li>\nmemory.provider = mnemosyne<\/code> in Hermes config,<\/li>\nFinal configuration<\/h2>\n
MNEMOSYNE_DATA_DIR=\/home\/jess\/.hermes\/mnemosyne\nMNEMOSYNE_EMBEDDING_API_URL=http:\/\/192.168.X.XXX:11434\/v1\nMNEMOSYNE_EMBEDDING_MODEL=nomic-embed-text:latest\nMNEMOSYNE_EMBEDDING_DIM=768\nMNEMOSYNE_LLM_ENABLED=false\n<\/pre>\n
Bottom line<\/h2>\n