SUMA Memory
Smart Universal Memory Architecture — the brain of QUAD. A knowledge graph with weighted relationships, RAG search, and explainable confidence scoring.
Why SUMA Exists
Every AI memory system stores text and finds “similar” text.
SUMA stores relationships with weights and finds the RIGHT context — not just similar context.
Others (Vector Search)
Convert question to numbers → find similar numbers → return text
No relationships, no importance, no paths, no explainability
SUMA (Graph + RAG)
Find entry points by meaning → traverse weighted graph → score paths → return ranked context with confidence
Relationships, importance, paths, explainability, cost control
The Nickname Test
Your team calls your CEO “Cookie” — a childhood nickname only insiders know. Someone asks your AI assistant: “What is Cookie working on?”
Vector Search (Mem0 / LangChain)
Converts “Cookie” to an embedding vector. Finds the closest matches by cosine similarity:
• “Best cookie recipes for team events”
• “HTTP cookie session management”
• “Cookie Monster Halloween costume ideas”
All wrong. No way to connect “Cookie” to a person.
SUMA (Graph + RAG)
Finds “Cookie” as a nickname node, then traverses the weighted relationship graph:
Cookie —nickname_of→ Ravi Kumar —role→ CEO
Ravi Kumar —working_on→ Series B Fundraise
Path strength: 0.85 • Confidence: High • 2 hops
Correct answer. Explainable path. Every hop traceable.
This is what “relationship-aware memory” means. Nicknames, aliases, indirect references — SUMA resolves them through graph traversal, not text matching.
Benchmark Results
Tested on LOCOMO (ACL 2024) — the same benchmark Mem0 uses in their published paper.
SUMA vs Mem0 — Head to Head
What Makes SUMA Different
A new class of memory architecture — patent-protected, built from first principles.
Relationship-Aware Retrieval
SUMA doesn't just find similar text — it follows relationships. A nickname resolves to a person. A project connects to its team. Context arrives with a traceable path, not a similarity score.
Explainable Confidence
Every answer comes with a confidence score and the reasoning behind it. You know WHY a result was returned — not just that it matched. No black box.
Memory Boundaries by Design
Personal memory and company memory are separate domains. Work data cannot bleed into personal. Health data stays in health. Boundaries enforced at retrieval, not just policy.
On-Device Ready
SUMA is designed to run sensitive queries entirely on-device. Your personal health memory never needs to leave your phone. This is an architectural guarantee — not a privacy policy.
How We Compare
Every memory platform has strengths. Here's an honest look at where each excels — and where SUMA goes further.
What Competitors Do Well
Mem0
- • Large open-source community (Apache 2.0)
- • Easy-to-use managed cloud API
- • User, org, & cross-app memory scoping
- • Self-hosted & cloud options
Zep
- • Session-based memory management
- • Built-in conversation summarization
- • BFS-based multi-hop graph traversal
- • Good for chatbot use cases
LangMem
- • Tight LangChain ecosystem integration
- • Flexible memory type abstractions
- • Open-source & self-hostable
- • Good for LangChain-based projects
Where SUMA Goes Further
| Capability | Mem0 | Zep | LangMem | SUMA |
|---|---|---|---|---|
| Vector / semantic search | Yes | Yes | Yes | Yes |
| Entity extraction | Yes | Yes | Partial | Yes |
| Weighted relationships | No | No | No | Yes |
| Multi-hop graph traversal | No | Yes | No | Yes |
| Path scoring (patented) | No | No | No | Yes |
| Explainable confidence | No | No | No | Yes |
| Precision control slider | No | No | No | Yes |
| LLM-free search option | No | Partial | No | Yes |
| Cross-application memory | Yes | No | No | Yes |
| Personal vs company separation | Yes | No | No | Yes |
| Self-hosted option | Yes | Partial | Yes | Yes |
Patent-Protected Capabilities
Multi-Hop Relationship Traversal
Follow chains of relationships to reach context that vector search can never find. Indirect connections resolved automatically.
Explainable Confidence
Every result comes with a score that tells you WHY it was returned — not just that it matched. No guessing.
Precision-Controlled Search
One control adjusts the tradeoff between breadth and precision. No tuning 10 parameters.
LLM-Optional Queries
Most queries resolve without calling an LLM at all. Costs drop by orders of magnitude at scale.
Context-Sensitive Depth
Search depth adapts to the strength of connections — close relationships go deeper automatically.
Memory Portability
Personal knowledge travels with the person. Company knowledge stays with the org. Enforced at the architecture level.
5 U.S. patents applied • 3 India patents in preparation • No competitor can replicate without licensing.
Memory That Travels With You
Every SQUAD application reads from and writes to the same SUMA graph. Your memory is shared, not siloed.
Start sprint for a client project → SUMA creates project context
QMS asks "What is the team working on?" → SUMA finds sprint via graph
"What did I work on today?" → Full day context from all apps
Personal Memory
Your skills and learnings. Travels when you change jobs via QUAD RESUME.
Company Memory
Customer data, IP, processes. Stays with the org. HIPAA/SOX/GDPR compliant.
Multi-Tenant
Every organization isolated by org_id. Same engine, zero data leakage.