Patent Pending • No competitor has this

SUMA Memory

Graph-Based Memory with Weighted Relationships. The brain of QUAD.

What Makes SUMA Different

Every AI memory system stores text and finds “similar” text.

SUMA stores relationships with weights and finds the RIGHT context — not just similar context.

Everyone Else (Vector Search)

User Question

  → Convert to numbers (embedding)

  → Find similar numbers in database

  → Return similar text

Fails for: relationships, importance, paths, explainability

SUMA (Weighted Graph)

User Question

  → Search graph nodes

  → Follow weighted edges

  → Calculate virtual weights across hops

  → Return ranked context with full path

Handles: relationships, importance, paths, explainability

The “Who is Cookie?” Test

Mem0 / LangChain

  1. 1. Convert “Cookie” to embedding
  2. 2. Find similar embeddings
  3. 3. Returns: “Cookie recipe” (0.89), “Cookie Monster” (0.87), “HTTP Cookie” (0.85)
  4. All wrong. Cost: $0.005

SUMA

  1. 1. Exact match on tags → no match
  2. 2. Fuzzy match (Levenshtein ≤2) → no match
  3. 3. LLM fallback: “Cookie = Kuvi, Lokesh's daughter”
  4. 4. Graph: Kuvi →(0.9)→ Lokesh →(0.95)→ Suman
  5. 5. Virtual weight: 0.9 × 0.95 × 0.85 = 0.727
  6. Correct answer. Cost: $0.001

SUMA vs The Competition

FeatureMem0ZepLangChainGoogleSUMA
Graph structureLabelsTemporalNoNoWeighted
Weighted relationshipsNoNoNoNoYes (0.0-1.0)
Virtual weight (multi-hop)NoNoNoNoYes (decay^d)
Path explainabilityNoNoNoNoYes
Hierarchical spheresNoNoNoTopicsYes (recursive)
LLM-free searchNoPartialNoNoYes (most)
Cloud-agnosticYesPartialYesNo (GCP)Yes
Cost per 1K queries~$5~$3~$2~$8~$0.10

The Patent Moat

India Patent 111 claims

Adaptive Weight-Based Context Depth

Uses numerical weights to determine how deep to search the graph. Higher weights = deeper traversal.

US Patent 63/956,81010 claims

Harmonic Mean Path Weighting

Calculates virtual weights across multi-hop paths: virtual_weight = max(path_weight × 0.85^distance).

India Patent 314 claims

Sphere-Aware Traversal

Hierarchical sphere structure scopes searches. Search relevant sphere first, expand only if needed.

Total: 50 patent claims across US + India. No competitor can replicate without licensing.

SUMA is a REST API

Any application can use SUMA. Nodes, relationships, weights — all accessible via HTTP. No vendor lock-in. No embedded SDK required.

POST
/nodes
Create a memory node with content, sphere, tags, weight
POST
/relationships
Connect two nodes with a weighted relationship (0.0-1.0)
GET
/search?q=...
Search nodes by keyword, fuzzy match, or graph traversal
GET
/graph/:id
Get node with all connections, paths, and virtual weights
GET
/stats
Total nodes, relationships, spheres, storage usage
DELETE
/nodes/:id
Soft delete with weight decay (preserves graph integrity)

The SUMMA Algorithm

Proprietary • Patented • No competitor has this

5-step search algorithm. 4 steps are free. 1 costs $0.001. Most queries never reach step 5.

1
Exact Match(40% of queries)

Search node tags and content for exact keyword match

$0.00
2
Fuzzy Match(25% of queries)

Levenshtein distance ≤ 2. Catches typos, nicknames, variations

$0.00
3
Graph Traversal(20% of queries)

BFS from matched nodes. Follow weighted edges. Apply weight decay per hop

$0.00
4
Virtual Weight(10% of queries)

Calculate multi-hop strength: harmonic mean × 0.85^distance

$0.00
5
LLM Fallback(5% of queries)

Only if steps 1-4 return nothing. Ask AI to interpret the query

$0.001

95% of queries resolved without any LLM call. Zero cost for algorithmic search.

Technical Specifications

Packagesuma-memory v0.3.0 (MIT License, PyPI)
StorageJSON (local) or PostgreSQL (multi-tenant)
Database2 tables: suma_nodes + suma_relationships (partitioned by org_id)
SearchKeyword (rapidfuzz), Semantic (sentence-transformers), Hybrid, Two-level RAG
GraphExplicit weighted relationships, BFS traversal, path finding, virtual weights
AdvancedDeduplication, node merging, lifecycle, soft delete with decay
CLI15+ commands
Tests30+ passing tests