How I Built a Wiki-Base (Context Engine) for the Latent Space Community
The file-vs-graph debate is a distraction. There is only the context window - and structure is what makes filling it efficiently possible.
I built a wiki-base (context engine) for the Latent Space community to put this to the test: SQLite backend, ingestion pipeline, vector + FTS5 search, MCP server, and a Discord bot. This video explains the thesis, shows the implementation, and argues why every individual and business should be building their own external context corpus.
0:00 Intro + why this matters
1:30 The file-vs-graph debate (it gets religious)
4:00 There is only the context window
7:30 Why structure enables reads, writes, and forcing functions
11:00 The year of sub-agents
13:00 SQLite is probably the best abstraction
15:00 The Latent Space Wiki-Base: schema (nodes + edges)
19:00 Skills and documentation for agents
21:00 Two interfaces: MCP server + Discord bot
24:00 Indexing: vector, FTS5, B-tree
26:30 Hub demo: searching + auto-created connections
28:00 Why descriptions need to be explicit (the enrichment pipeline)
30:30 How to get involved
---
Links:
• Latent Space community: https://www.latent.space/
Build your own context corpus:
• RA-H (free Mac app): https://ra-h.app
• RA-H (open source): https://github.com/bradwmorris/ra-h_os/
#contextengineering #pkm #sqlite #aiagents #latentspace