memgraph — Scribe Mem structured memory service

Memgraph is Robonix's structured CKG memory system (formerly Scribe Mem): it consumes LogRecord entries from Scribe Log and builds a causal knowledge graph (CKG) with tag-indexed, vector-searchable memory nodes.

Runs in parallel with memsearch under the shared robonix/service/memory/ namespace — Pilot discovers both backends and the LLM chooses between them.

Capability surface

Contract	Behaviour
`robonix/service/memory/remember`	Write a MemoryNode (tag → vector → graph)
`robonix/service/memory/hybrid_search`	Tag filter → BM25+Embedding hybrid → causal filter
`robonix/service/memory/promote`	Promote short-term → long-term nodes

Architecture

LogRecord ──[remember]──► MemoryNode
                              ├── TagIndex (inverted, O(1) filter)
                              ├── VectorStore (all-MiniLM-L6-v2 d=384, BM25+Cosine)
                              └── GraphStore (JSON persistence, causal edges)

Run standalone

cd services/memory
PYTHONPATH=. python3 -c "
import asyncio
from memory_service.service import MemoryService
svc = MemoryService()
asyncio.run(svc.init())
print(f'Nodes: {svc.graph.count()}')
"

Configuration

EMBEDDING_MODEL_PATH — path to all-MiniLM-L6-v2 model directory (optional; otherwise sentence-transformers uses its standard model cache)
Spatial records containing objects must provide their actual coordinate frame in origin; missing origins are rejected rather than labelled world.
Falls back to deterministic hash embeddings when model unavailable

Status

Phase 1 — core remember + search + compact. Ready for Demo 1 (object spatiotemporal backtracking) and Demo 2 (task history backtracking).