// Build or rebuild the knowledge graph from existing memories using TF-IDF entity extraction and Leiden clustering. Use when search results seem poor, after bulk imports, or to optimize performance. Automatically discovers relationships between memories and creates topic clusters.
| name | slm-build-graph |
| description | Build or rebuild the knowledge graph from existing memories using TF-IDF entity extraction and Leiden clustering. Use when search results seem poor, after bulk imports, or to optimize performance. Automatically discovers relationships between memories and creates topic clusters. |
| version | 3.4.22 |
| license | AGPL-3.0-or-later |
| compatibility | Requires SuperLocalMemory V2 installed at ~/.claude-memory/, optional dependencies: python-igraph, leidenalg |
| attribution | {"creator":"Varun Pratap Bhardwaj","role":"Solution Architect & Original Creator","project":"SuperLocalMemory V2"} |
Build or rebuild the knowledge graph from existing memories to improve search quality and discover hidden relationships.
slm build-graph [--force] [--clustering]
$ slm build-graph
Output:
🔄 Building Knowledge Graph...
Phase 1: Entity Extraction
Scanning 1,247 memories...
Extracted 892 unique entities
Created 892 graph nodes
✓ Complete (3.2s)
Phase 2: Relationship Discovery
Computing similarity scores...
Created 3,456 edges (relationships)
Avg edges per node: 3.9
✓ Complete (5.1s)
Phase 3: Optimization
Indexing graph structure...
Pruning weak edges (score < 0.3)...
Final edge count: 2,134
✓ Complete (1.2s)
✅ Knowledge graph built successfully!
Graph Statistics:
Nodes: 892
Edges: 2,134
Density: 0.27%
Largest Component: 856 nodes (96%)
Next: Use `slm recall` to see improved search results
$ slm build-graph --force
Rebuilds from scratch (deletes existing graph first)
Use when:
$ slm build-graph --clustering
Requires optional dependencies:
pip3 install python-igraph leidenalg
Additional output:
Phase 4: Topic Clustering (Leiden)
Detecting communities...
Found 47 clusters
Largest cluster: 89 memories
Smallest cluster: 3 memories
Modularity score: 0.82 (excellent)
✓ Complete (2.3s)
Discovered Clusters:
Cluster 1 (89 memories): "Authentication & Security"
Top entities: JWT, OAuth, tokens, auth, security
Cluster 2 (76 memories): "Database & PostgreSQL"
Top entities: PostgreSQL, database, SQL, queries, indexes
Cluster 3 (54 memories): "React & Frontend"
Top entities: React, hooks, components, state, props
...
| Argument | Description | When to Use |
|---|---|---|
--force | Delete existing graph and rebuild | Corruption, fresh start |
--clustering | Run topic clustering | Want to discover topic groups |
--verbose | Show detailed progress | Debugging, understanding process |
--dry-run | Preview without saving | Testing, analysis |
Entities extracted from memories:
Node properties:
Relationships between entities:
Edge properties:
Topic groups discovered:
| Memory Count | Build Time | Notes |
|---|---|---|
| 100 | ~1s | Instant |
| 1,000 | ~10s | Fast |
| 10,000 | ~2min | Acceptable |
| 50,000+ | ~15min | Plan accordingly |
With clustering (add ~50%):
Factors affecting speed:
# Add new memories
slm remember "New content..." --tags new
# Incremental graph update (fast)
slm build-graph # Only processes new memories
# Force full rebuild (slower, thorough)
slm build-graph --force
# Check graph stats before
slm status | grep "Knowledge Graph"
# Build graph
slm build-graph --verbose
# Check stats after
slm status | grep "Knowledge Graph"
Weekly rebuild (cron job):
#!/bin/bash
# Every Sunday at 3 AM
echo "$(date): Starting graph rebuild"
slm build-graph --clustering >> /var/log/slm-build.log 2>&1
echo "$(date): Graph rebuild complete"
Post-import hook:
#!/bin/bash
# After bulk import
memories_added=$1
if [ "$memories_added" -gt 50 ]; then
echo "Large import detected, rebuilding graph..."
slm build-graph
fi
# Build with clustering
slm build-graph --clustering
# Check discovered clusters
slm status --verbose | grep -A 20 "Topic Clusters"
# Search within specific cluster
slm recall "FastAPI" --cluster "Backend & APIs"
Cause: Not enough RAM for large graph
Solution:
# Build in chunks (process fewer memories at once)
slm build-graph --chunk-size 1000
# Or increase system memory
# Or archive old memories
Cause: Optional dependencies not installed
Solution:
pip3 install python-igraph leidenalg
# Verify
python3 -c "import igraph; import leidenalg"
# Try again
slm build-graph --clustering
Causes:
Solutions:
# Show progress
slm build-graph --verbose
# Skip clustering (faster)
slm build-graph # No --clustering flag
# Check disk space
df -h ~/.claude-memory/
Cause: Stale graph or poor similarity threshold
Solution:
# Force complete rebuild
slm build-graph --force
# Adjust similarity threshold (advanced)
slm build-graph --min-similarity 0.4 # Default: 0.3
Total unique entities found
Why matters: More nodes = richer semantic understanding
Total relationships discovered
Why matters: More edges = better search via relationships
How connected the graph is
Size of biggest connected subgraph
Why matters: Smaller component = isolated knowledge islands
How well-defined clusters are
Why matters: Higher = clearer topic separation
Before graph build:
After graph build:
Example:
Query: "authentication"
Before:
- Direct matches only (JWT, auth, login)
After:
- Direct matches (JWT, auth, login)
- + Related concepts (security, tokens, OAuth)
- + Connected memories (API design, user management)
Shows updated graph statistics
Each profile has separate graph
slm recall - Search uses the graphslm status - Check graph statsslm remember - Add memories (triggers incremental update)Created by: Varun Pratap Bhardwaj (Solution Architect) Project: SuperLocalMemory V2 License: MIT (see LICENSE) Repository: https://github.com/varun369/SuperLocalMemoryV2
Open source doesn't mean removing credit. Attribution must be preserved per MIT License terms.
Build or rebuild the knowledge graph from existing memories using TF-IDF entity extraction and Leiden clustering. Use when search results seem poor, after bulk imports, or to optimize performance. Automatically discovers relationships between memories and creates topic clusters.
List most recent memories in chronological order. Use when the user wants to see what was recently saved, review recent conversations, check what they worked on today, or browse memory history. Shows memories sorted by creation time (newest first).
Search SuperLocalMemory for relevant facts, decisions, and past context. Use when the user asks to recall, search, find, or retrieve stored information. Invokes 5-channel retrieval with LightGBM reranking via MCP.
Save content to SuperLocalMemory with intelligent indexing and knowledge graph integration. Use when the user wants to remember information, save context, store coding decisions, or persist knowledge for future sessions. Automatically indexes, graphs, and learns patterns.
Check SuperLocalMemory system status, health, and statistics. Use when the user wants to know memory count, graph stats, patterns learned, database health, or system diagnostics. Shows comprehensive system health dashboard.
Switch between memory profiles for context isolation and management. Use when the user wants to change profile context, separate work/personal memories, or manage multiple independent memory spaces. Each profile has its own database, graph, and patterns.