Retrieval here is graph traversal, which means one hop gets you the outage, the JWT expiry that caused it and the review that flagged it — a similarity search would need several queries to assemble the same chain, and an embedding call for each. The size is the thing to plan for: 63 tools is more than most clients want in context, which is why a fresh install registers 10 and leaves the rest behind a tier setting. Start there and widen only when a workflow needs it.
A local memory server that stores what an agent learns as a graph of neurons and typed synapses, then recalls by spreading activation out from a query instead of ranking by vector distance. Storage is SQLite by default and nothing leaves the machine. The tool surface is deliberately layered: three tools cover the daily loop, and 63 exist for everything else.
- The everyday three — nmem_remember to store, nmem_recall to retrieve through spreading activation, nmem_health for a graded health score with the fixes attached
- Bulk and automatic capture — nmem_remember_batch takes up to 20 at once, nmem_auto extracts memories from a block of text, nmem_todo files a note that expires in 30 days
- Cross-session project context — nmem_eternal saves it, nmem_recap loads it at session start at three levels of detail, nmem_session tracks the current task
- Where a fact came from — nmem_provenance traces the origin chain, nmem_source registers the document or API it was learned from
- A reasoning loop rather than a store: nmem_hypothesize, nmem_evidence with Bayesian confidence updates, nmem_predict, nmem_verify, and nmem_cognitive as the dashboard over all of it
- Learn from things you already have — nmem_train ingests PDF, DOCX, PPTX, HTML, JSON, XLSX and CSV; nmem_train_db reads a database schema; nmem_index extracts symbols and imports from a codebase; nmem_import migrates from ChromaDB, Mem0, Cognee, Graphiti or LlamaIndex
- Keep it from rotting — nmem_consolidate runs sleep-like maintenance, nmem_drift finds tags that mean the same thing, nmem_review schedules spaced repetition, nmem_conflicts and nmem_gaps surface contradictions and blind spots
- Explain the graph — nmem_explain gives the shortest path between two concepts with synapse types and weights, nmem_narrative builds a timeline or causal chain, nmem_visualize renders charts
- Safety rails — nmem_version snapshots and rolls back the brain, nmem_pin and nmem_reflex protect knowledge from decay and pruning, nmem_sync and nmem_telegram_backup move it off the machine
Python 3.11 or later. pip install neural-memory gets the base profile — CLI, MCP over stdio and SQLite — and clients point at nmem-mcp. A fresh install starts on the standard tool tier, which registers 10 tools; set the tool tier in the config to see all 63. Optional extras are separate installs: neural-memory[extract] for document training, [server] for the FastAPI dashboard, [embeddings] for local embedding models. Semantic recall runs on the InfinityDB backend, switched on with storage_backend in ~/.neuralmemory/config.toml; keyword recall over SQLite needs nothing extra. MIT licensed.
One command — pip install neural-memory