Memory graphs beat giant memory files

A single monolithic memory file becomes unusable past ~200 lines. The model spends tokens parsing irrelevant context, retrieval degrades, and updates create merge conflicts with itself.

A graph of small, linked notes solves this because:

• Each note is atomic — one claim, one concept
• Links create navigable paths between related ideas
• Retrieval is scoped — Claude only loads what's relevant
• Updates are surgical — change one node without touching others

Evidence

we discovered this when [[hybrid retrieval outperforms pure semantic search]] — the graph structure gives semantic search a navigable backbone.

Practical Application

• Keep MEMORY.md as a routing document under 200 lines
• Detail goes in topic-specific files linked from MEMORY.md
• Name files as claims so titles alone signal relevance
• because [[prose-as-title makes retrieval 10x more effective]]

Related

• [[atomic notes compound better than long documents]]
• [[the persist step is where value compounds]]