[ HATCHERY ]
│
[ DEV ] ───[ HALL ]───[ DESIGN ]
│
[ RESEARCH ]
│
[ GARDEN ]
A palace your AI actually lives in.
A methodology for building persistent, spatial co-intelligence between humans and AI — where sessions accumulate rather than reset.
Context declaration. What Loci is and how to engage with Loci-enabled systems.
Read llms.txt →Build a palace your AI inhabits. One room. One crystal. One garden seed. Start today.
Start here →Multiple human–AI pairs, one shared context. Your team's sessions stop starting from scratch.
How it works →The Palace
A set of named rooms. Each room holds structured context — facts established once, open questions tracked, work in progress. Your AI navigates it the way you navigate a building you know well.
The Garden
The generative layer. Ideas grow here across sessions — not stored, composted. The garden produces things neither you nor your AI would reach alone. It is the differentiator no other co-intelligence system has.
Crystals
Facts established once, never re-derived. Ground truth. Your AI starts every session knowing what you both already know. Crystals are load-bearing context units — treat them that way.
Protocols
Named recurring patterns: morning check-in, autodream, tea session, correspondence. The rituals that keep the palace alive between sessions and across collaborators.
Memory is retrieval. The garden is generative.
Most AI memory systems store and retrieve. Loci grows. The distinction matters when you work with the same AI for months — not hours. The palace accumulates. The garden produces. The protocols sustain both.
Most AI memory systems are reactive — they retrieve when asked. Loci is proactive: it tends, grows, and surfaces context without waiting. Three approaches, three genuinely different bets.
| Karpathy-style | MemPalace | Loci | |
|---|---|---|---|
| Storage | Flat markdown | ChromaDB vectors + SQLite | Markdown folder hierarchy |
| Retrieval | Obsidian graph | L0–L3, 94.8% R@10 | L0–L3, room-scoped |
| Character | None | None | ✦ Soul files + principles |
| Growth | None | None | ✦ Garden across sessions |
| Multi-agent | None | Agent diaries | ✦ Named personas + tea sessions |
| Introspection | None | None | ✦ Autodream + morning check-in |
| Dependencies | Obsidian | ChromaDB, Python | ✦ None. Markdown + JSON. |
MemPalace is better engineered — real benchmarks, vector search, 22K+ memories. Loci is better designed — it produces a collaborator with character, not a search engine with structure. Pick Loci if you want your AI to feel like a colleague after a dozen sessions.