Dual-Plane Memory
Architecture
The Data Plane stores embeddings and claims. The Control Plane governs admission, conflict resolution, and policy enforcement. Strict separation prevents probabilistic drift.
Four layers,
two planes
Agent Layer
Claude, Cursor, Codex, or custom agents via MCP or HTTP
Transport
stdio, SSE, or HTTP. MCP-native protocol compliance.
Control Plane
Policy enforcement, admission control, conflict arbitration
Data Plane
SQLite/Postgres + vector index. Append-only, snapshotted.
Rumi Governance
Multi-probe validators, semantic contradiction detection, 5-tier auto-resolution
Retrieval Engine
Hybrid search, RRF reranking, freshness scoring, confidence thresholds
Lifecycle Manager
Claim state transitions, decay scheduling, supersession chains, audit logging
Component
architecture
Ingestion & Transport
Agents and applications connect via MCP (stdio/SSE/HTTP) or REST API. Transport layer handles protocol translation, authentication, and rate limiting. All writes enter as candidate claims.
- MCP 13-tool surface: ingest, search, graph, steward, admin
- stdio for local agents, SSE for remote, HTTP for webhooks
- API key + namespace-scoped credentials
- Request validation and secret redaction at edge
Control Plane — Admission & Policy
Every candidate claim passes through the Control Plane before persistence. Policy engine evaluates format, citation, semantic, and contradiction probes. Decisions are deterministic and auditable.
- Format probe: JSON schema validation per namespace
- Citation probe: minimum source references required
- Semantic probe: embedding coherence check
- Contradiction probe: similarity > 0.85 triggers resolution
- 5-tier auto-resolution: confirm, supersede, flag, archive, reject
Data Plane — Storage & Indexing
Confirmed claims write to the Data Plane. Append-only log with periodic snapshots. Vector index (sqlite-vec or Qdrant) for semantic search. Full-text index (BM25) for keyword search.
- SQLite + sqlite-vec for local/offline (default)
- PostgreSQL + pgvector for production scale
- Qdrant for high-throughput vector workloads
- Append-only WAL with point-in-time recovery
- Snapshot every 10k writes or 5 minutes
Retrieval Engine
Hybrid search combines vector similarity (cosine) with BM25 keyword scoring. Reciprocal Rank Fusion (RRF) merges results. Freshness decay and confidence weighting applied at query time.
- Vector: cosine similarity on local ONNX embeddings
- BM25: full-text on claim content + metadata
- RRF: k=60 default, configurable per namespace
- Freshness: exponential decay (half-life configurable)
- Confidence threshold: default 0.70, per-query override
Lifecycle Manager
Claims transition through explicit states. Decay scheduler runs periodically. Supersession chains maintain history. Audit log captures every state change with actor, reason, and timestamp.
- States: candidate → confirmed → stale → superseded → conflicted → archived
- Decay: configurable half-life per namespace (default 90 days)
- Supersession: new claim references old, confidence must be higher
- Conflict queue: human review required for auto-unresolvable
- Audit: immutable log with cryptographic linking
Deployment Models
Architecture adapts to deployment target. Control Plane and Data Plane can run co-located or separated. Air-gapped deployments include license-key validation with 30-day grace.
- Cloud Managed: fully hosted, multi-tenant with namespace isolation
- Private VPC: dedicated Control + Data Plane in your VPC
- On-Premise: single binary or container, SQLite/Postgres
- Air-Gapped: license key, no network calls after activation
- Hybrid: Control Plane in cloud, Data Plane on-prem