Knowledge Graph Patterns

Engram’s knowledge graph connects memories with typed, directed edges. This guide covers patterns for building graphs that make recall more powerful.

Edge Types

Relation	Meaning	Example
relates_to	General association	Auth middleware ↔ JWT library choice
supersedes	Replaces outdated knowledge	New API design supersedes old one
contradicts	Conflicts with	Two conflicting approaches found
part_of	Component of a larger concept	Detail memory is part of hub memory
derived_from	Conclusion drawn from source	Summary derived from debugging session
summarized_by	Condensed version	Cluster of memories summarized by one

Hub-and-Spoke Pattern

For complex topics, create a brief hub memory first, then store details as children:

{"tool": "remember", "content": "Auth system architecture overview: JWT + refresh tokens, httpOnly cookies, 15-min expiry"}

Then add detail memories with parent_id to automatically create part_of edges:

{"tool": "remember", "content": "JWT signing uses RS256 with rotating keys stored in AWS KMS", "parent_id": 42}

Supersession Chain

When decisions evolve, create supersedes edges rather than deleting old memories:

{"tool": "connect", "from": 55, "to": 42, "relation": "supersedes"}

This preserves history — you can trace why decisions changed.

Team Graph Patterns

In team contexts, the hive mind consolidation process uses these patterns automatically:

• Deduplication: When two engineers remember the same thing, memories are merged
• Contradiction detection: Conflicting knowledge gets contradicts edges, flagging it for resolution
• Cross-team synthesis: Insights spanning domains are connected with relates_to edges

Using Depth in Recall

Set depth: 1 (up to 3) when recalling to follow edges and surface connected knowledge:

{"tool": "recall", "query": "authentication", "depth": 1}

This returns not just matching memories, but their graph neighbors — surfacing related context that flat search would miss.