Glossary¶
Domain terms used across HydraMem. Use these definitions to avoid semantic drift.
| Term | Definition |
|---|---|
| SR-MKG | Structured Relation – Mutual Knowledge Graph. A pure-Python, topological scorer that ranks/filters candidate relations using graph structure and heuristic weights — no LLM. Stage 1 of verification. |
| VoG | Verification of Groundedness. An LLM step that checks whether a candidate is grounded in retrieved evidence. Stage 2 of verification; also used for the chunk path. |
| Chunk prefilter | The vector-similarity (cosine) filter applied to chunks inside hydra_search. It is not SR-MKG; only relations go through SR-MKG. |
| Night Gardener | The autonomous offline cycle that runs in three phases — infer → verify → prune — to refine the graph. Honest contract: emits zero relations when there is no real evidence. |
| Fast path | Low-latency retrieval with no LLM: vector ANN + entity term match + 1-hop expansion. Used by priming_context. |
| Full path | High-precision retrieval: vector ANN + multi-hop graph traversal + SR-MKG + VoG. Used by hydra_search. |
priming_context |
Fast-path tool returning top-k chunks + neighbouring entity names for the agent to expand. |
hydra_search |
Full-path tool returning verified context with metadata. |
expand_context |
Multi-hop graph traversal from known entity IDs. |
trace_path |
Shortest path between two entities (NetworkX shortest_path). |
| MENTIONS | The edge type linking a chunk to an entity it references. |
| Hyper-relational qualifier | Extra key/value context attached to a relation (subject–predicate–object + qualifiers), making relations first-class hyper-edges. |
| PPR | Personalized PageRank retrieval (HippoRAG-style), in hydramem/ppr.py. |
| CRDT | Conflict-free Replicated Data Type, used by the garden (hydramem/garden/crdt.py) for mergeable state. |
| LightGNN pruning | Optional GNN-based spurious-edge detection (hydramem/gnn_prune.py); heuristic fallback; auto-skips on graphs larger than HYDRAMEM_GNN_MAX_NODES. |
| Honesty contract | The project rule that no feature/metric is described as working unless it measurably works; the dashboard must reflect reality. |
| Shadow estimator | The naive-RAG baseline used to estimate token savings (hydramem/telemetry/shadow.py); the source of the "tokens saved" metric. |
| Grafeo / LadybugDB | The persistent graph backend (Grafeo = Rust core via PyO3; LadybugDB = a Kuzu fork). On Python 3.11 the graph falls back to a NetworkX persistent store. |
| LanceDB | The persistent vector store for chunks + embeddings. |
| fastembed | The default lightweight ONNX embedder (~80 MB, no torch); falls back to sentence-transformers when configured. |
| MCP | Model Context Protocol — the agent-facing tool protocol served by FastMCP in hydramem/server.py. |
| Project namespacing | Per-project isolation of stored data so multiple knowledge bases coexist in one deployment. |
min_repeat_count |
Night Gardener config gating how many times a session snapshot must repeat before it is considered for inference (default 2). |
| Calibration | Per-project tuning of SR-MKG/verification thresholds (tests/test_calibration.py). |