Neuromcp

Get all memories in a specific cluster, ordered by distance from centroid.

Get details of a specific episode including memory count.

Import memories from JSONL or JSON data. Deduplicates by content hash.

Initialize spaced repetition for important memories that do not have a review schedule yet.

List all clusters in a namespace with their labels and memory counts.

List episodes in a namespace. Shows memory count per episode.

Log a retrieval event so future searches can learn which memories are actually helpful. Call right after search_memory. Pass retrieved_ids (all top-K), cited_ids (those actually used in the answer, if known), and an outcome label (helpful/neutral/harmful). Usefulness scores are applied as a prior in future rankings.

Get statistics about stored memories: counts, categories, trust levels, importance, and database size.

Track how knowledge about a topic evolved over time. Follows supersession chains and shows full revision history.

Traverse the knowledge graph starting from an entity. Returns connected nodes and edges up to max_depth hops. Supports temporal queries. Without entity_id/entity_name it returns an OVERVIEW: the top-N entities of the namespace ranked by number of relations (contradicts edges not counted), with the relations among them; internal memory_proxy entities that only back contradicts edges are excluded from the overview.

Answer a question FROM memory: runs hybrid retrieval, then returns a synthesized, CITED extractive answer (every sentence traces to a stored memory id) PLUS an explicit gap-analysis — what memory does NOT cover and the freshness boundary (stale_since). Returns status "not_in_memory" instead of fabricating when nothing matches. Deterministic, no LLM. Prefer this over search_memory when you need a grounded answer rather than raw chunks.

Recall memories by ID, namespace, category, or tags without semantic search.

Register or update an agent profile with expertise topics. Auto-counts memories.

Record a review of a memory. Quality 0-5 (SM-2 scale: 0=forgot, 5=perfect). Updates interval and schedules next review.

Get memories due for spaced repetition review. Returns overdue memories ordered by urgency + importance.

Unified search across both extracted memories and verbatim text. Returns results with source labels and per-source quotas.

Search atomic claims extracted from memories. Claims are verifiable facts with subject-predicate-object structure.

Search memories using hybrid vector + full-text search with RRF ranking, graph boost, and cognitive priming. Supports temporal queries (valid_at) to find what was true at a specific time.

Search raw verbatim text using full-text search. Returns exact matches from stored conversations. Use for literal recall.

Start a new episode (session/task context). Memories stored with this episode_id will be grouped together. Use to track what happened during a session.

Store a new memory with semantic deduplication, contradiction detection, surprise scoring, and entity extraction. Returns the memory ID, contradictions found (resolution supersede/coexist are claim-backed and recorded as graph edges; flag is heuristic-only and reported here but never stored in the graph), surprise score, and extracted entities.

Store raw conversation text verbatim — no summarization, no consolidation, never pruned. Use for exact recall of what was said.

Generate an extractive summary of a cluster. Selects the most central, representative sentences from cluster memories using embedding centrality.

Generate an extractive summary of an episode. Selects the most central, representative sentences from episode memories.

Copy memories from source to target namespace. Optionally adapts content by stripping project-specific references.

Recalculate adaptive importance for all memories in a namespace. Boosts frequently accessed, recently relevant, and graph-central memories. Run after clustering and PageRank for best results.

List memories ranked by observed usefulness (helpful vs harmful citation ratio). Use to inspect which memories the agent actually leans on, or to pick high-confidence source material for reflection.

Get statistics about verbatim storage: total entries, size, and distribution by namespace.

Generate a structured briefing from the wiki: active projects, unprocessed sources, recent changelog entries, stale pages needing attention. Use for morning status checks.

Read a file from the wiki raw-sources/ directory and extract structured metadata (title, type, key concepts, related pages). Returns analysis the LLM uses to decide which wiki pages to create or update. Does NOT write pages itself.

Scan all wiki pages and check for health issues: missing frontmatter, pages not in index.md, stale pages (>30 days), thin pages (<5 lines), unprocessed raw sources, broken related links. Writes lint-report.md.