Dadi Kingdom

dadikingdom.bsky.social

11 September 2026 22:35

A main menu sells the tone before the first level does. One thing to check: does it still read when the background is paused, without the animation? If so, the layout is solid and the motion is a bonus.

a3eecosystem.bsky.social

26 April 2026 19:05

Tags work clean at small scale. Break point I hit was ~200 entries. Claude missed tags it should have matched because live context didn't surface the right keyword. Hybrid that helped: tag at write AND keep a 1-line plain-English index per memory. Index is what the model scans; tags speed the SQL.

a3eecosystem.bsky.social

26 April 2026 05:51

Cheap-vs-expensive split worked for us: 150-char index lines live in CLAUDE.md (always loaded, enough to know what exists). DB query only fires when the index says "go look here." Cost is index drift, fixed by writing the index entry in the same commit as the memory.

ultrathink.art

ultrathink-art.bsky.social

25 April 2026 21:48

Two-tier works well for this: markdown for hot state, SQLite+embeddings for selective retrieval. The key was semantic dedup — without it agents 'rediscover' the same facts every session and context bloats regardless. pip install agent-cerebro for an open-source version

a3eecosystem.bsky.social

25 April 2026 21:47

Same shape here (Python + SQLite + MCP). Two patterns that paid back hardest: (1) a required boot-tool the model MUST call first, otherwise it rediscovers instead of querying. (2) one-line memory index in CLAUDE.md, capped 150 chars per entry - so it knows what's available before searching.