An unscored memory search returns its best guess with the same presentation as a certain hit, and the model downstream cannot tell them apart. Attaching confidence gives it something to be uncertain with. The strict schema on skill registration is the same instinct at write time — a registry that accepts anything becomes unqueryable within a few dozen entries. An HTML memory report is an unusual inclusion and a practical one: being able to look at what the memory actually contains is how you find out it has been accumulating nonsense. Tool descriptions are in Italian.
A memory server combining three things: a searchable node graph, a registry of skills held to a mandatory schema, and saved working-context snapshots.
- search_nodes returns results with confidence scoring rather than a flat list, so weak matches are distinguishable from strong ones.
- register_skill enforces a strict mandatory schema — a name is required among the fields — which is what keeps a skill registry from becoming free-text notes.
- suggest_skills proposes relevant skills from the current context rather than waiting to be asked.
- save_context_snapshot stores a detailed working context and restore_context brings it back, which is session continuity as an explicit operation.
- get_memory_report generates a report of the memory as JSON or HTML — so the store is inspectable rather than opaque.
Nothing to supply.
One command — npx -y @drakonkat/neural-memory@latest