# Automatic Context Extraction
When creating relationships between memories, MemoryGraph automatically extracts structured information from natural language context. This enables powerful filtering and search while maintaining full backward compatibility.
## How It Works
Write natural language when creating relationships:
```python
create_relationship(
from_memory_id="mem_123",
to_memory_id="mem_456",
relationship_type="SOLVES",
context="partially implements auth module, only works in production"
)
```
MemoryGraph automatically extracts:
```json
{
"text": "partially implements auth module, only works in production",
"scope": "partial",
"components": ["auth module"],
"conditions": ["production"],
"evidence": [],
"temporal": null,
"exceptions": []
}
```
## Structured Query Support
Search relationships by their extracted context structure:
```json
{
"tool": "search_relationships_by_context",
"scope": "partial",
"conditions": ["production"],
"limit": 10
}
```
### Available Filters
- **scope**: `"partial"`, `"full"`, or `"conditional"` - Filter by implementation completeness
- **conditions**: `["production", "Redis enabled"]` - Filter by when/where the relationship applies (OR logic)
- **has_evidence**: `true` / `false` - Filter by whether the relationship has verification
- **evidence**: `["integration tests", "unit tests"]` - Filter by specific evidence types (OR logic)
- **components**: `["auth", "Redis"]` - Filter by mentioned components or modules (OR logic)
- **temporal**: `"v2.1.0"` - Filter by version or time information
- **limit**: `20` - Maximum number of results (default: 20)
### Example Queries
Find all partial implementations:
```json
{
"tool": "search_relationships_by_context",
"scope": "partial"
}
```
Find production-only relationships verified by tests:
```json
{
"tool": "search_relationships_by_context",
"conditions": ["production"],
"has_evidence": true
}
```
Find all auth-related relationships:
```json
{
"tool": "search_relationships_by_context",
"components": ["auth"]
}
```
Combined filters (uses AND logic between different filter types):
```json
{
"tool": "search_relationships_by_context",
"scope": "partial",
"conditions": ["production"],
"components": ["auth"],
"has_evidence": true
}
```
## Extracted Fields
The system recognizes and extracts these categories:
- **scope**: Completeness indicators (partial, full, conditional)
- **components**: Mentioned modules, systems, or services
- **conditions**: When/if/requires patterns (environments, dependencies)
- **evidence**: Verification mentions (tested by, verified by, observed in)
- **temporal**: Version numbers, dates, or time-based information
- **exceptions**: Exclusions or limitations (except, excluding, but not)
## Examples
```python
# Scope detection
"fully supports user authentication" → scope: "full"
"limited to admin users" → scope: "partial"
# Conditions
"only works in production environment" → conditions: ["production"]
"requires Redis to be available" → conditions: ["Redis to be available"]
# Evidence
"verified by integration tests" → evidence: ["integration tests"]
"tested by QA team" → evidence: ["QA team"]
# Temporal
"implemented in v2.1.0" → temporal: "v2.1.0"
"since March 2024" → temporal: "March 2024"
# Exceptions
"supports all formats except XML" → exceptions: ["XML"]
"works without authentication" → exceptions: ["authentication"]
```
## Token Trade-off
Structure adds approximately 8-13 tokens per relationship context for storage:
- **Original text only**: ~12 tokens
- **Structured format**: ~20-25 tokens
The benefit is queryability and relationship intelligence, not token reduction. The original text is always preserved for readability.
## Backward Compatibility
All existing relationships with free-text contexts continue to work. The system automatically:
1. Tries parsing as JSON (new structured format)
2. Falls back to pattern extraction (legacy free text)
3. Returns empty structure if context is null/empty
No migration or changes required to existing data.
