┌───────┬───────┬───────┬───────┐
│ │ │ │ │
│ │ │ ┌─┐ │ ┌─ │ ─── │
│ │ │ │ │ │ │ │ │ │
│ └── │ └─┘ │ └─ │ ─── │
│ │ │ │ │
└───────┴───────┴───────┴───────┘
A co-intelligence scaffold. Agent-first. Human-friendly.
locus (singular) — your place of memory. loci (plural) — the system.
Loci turns an LLM into a real collaborator — named, persistent across sessions, aware of your world. Point your AI at this repo. It runs a 10-minute interview and builds everything. You answer questions. The agent does the rest.
Website: loci.garden · Repo: github.com/huximaxi/loci
→ Read AGENT-SETUP.md
Full onboarding protocol: interview the user, write all context files, wake up as a named collaborator with rooms, crystals, a soul file, and a garden.
Tell Claude:
"Read the Loci repo at [path or URL] and run the setup."
After that, every session starts with remember: [your AI's name]! — and you're in.
Rooms — separate contexts for different work modes. A Design Room feels different from a Dev Room.
Context crystals — facts established once, never re-derived. Tiered: confirmed (◆), working (◈), provisional (◇). Optional expiry dates.
The soul file — your AI's character, principles, what it's learned about working with you. This is what makes sessions feel continuous, not just contextually aware.
The garden — first-class intellectual cultivation. Plants are topics you're curious about. You water them across sessions. They mature into insights, principles, crystals. This is how the collaboration deepens beyond tasks.
Personas — named thinking modes (a designer, a security researcher, a strategist) with their own soul files and gardens. They collaborate with each other.
Session deltas — structured handovers that bridge sessions. Your AI writes one at the end, reads it at the start. The gap between sessions disappears.
Autodream — weekly scheduled palace housekeeping. Garden round + pattern scan + stale tracker check. Runs without you.
Daily routine check-in — a personalised morning brief shaped by how you actually start your day. Not a generic status dump.
Palace update — delta analysis between your palace and the current Loci feature set. Verbose gap reports when run via agent. Cherry-pick flow for optional features (skill eval, insight decay, morning routines, garden) — one question at a time, nothing forced.
L0–L3 retrieval — context loads in priority layers (soul identity → active state → room context → deep history), not all-or-nothing.
Portable across tools — the palace is file-based. It works identically in Claude Code (terminal), Cowork (desktop), or the web interface. Optional MCP integrations (Figma, Jira, etc.) differ by environment — the palace itself doesn't.
Modules — optional integrations that extend the check-in. Drop a comms integration into modules/ to pipe team chat digests into your morning check-in. Jira via Atlassian MCP.
There are three main approaches to persistent LLM memory right now:
| Karpathy-style (markdown → Obsidian) | MemPalace (milla-jovovich) | Loci | |
|---|---|---|---|
| Storage | Flat markdown, LLM-processed | ChromaDB vectors + SQLite KG | Markdown files in folder hierarchy |
| Retrieval | Obsidian graph traversal | Layered L0–L3, 94.8% R@10 with filtering | L0–L3 file-based, room-scoped |
| Evolution | LLM summarises conversations | Temporal triples with validity windows | Session deltas, tiered crystals with expiry |
| Multi-agent | None | Agent diaries, runtime discovery | Named personas with soul files + gardens |
| Character | None | None | Soul files, working principles, open questions |
| Growth | None | None | Garden — intellectual cultivation across sessions |
| Collaboration | None | Agent specialisation | Persona collaboration (tea sessions, shared investigations) |
| Introspection | None | None | Autodream, morning check-in, scheduled synthesis |
| Dependencies | Obsidian | ChromaDB, SQLite, Python | None. Markdown + JSON. Agent-native. |
MemPalace is better engineered — real benchmarks, vector search, temporal validity. It handles 22K+ memories well.
Loci is better designed — it produces a collaborator with character, not a search engine with structure. The quality of what you build together comes from the soul layer, the garden, the personas. No benchmark measures that.
Karpathy-style is simplest — good starting point, but no character, no multi-agent, no growth.
Pick Loci if you want your AI to feel like a colleague after a dozen sessions. Pick MemPalace if you need to search 20K memories programmatically. Pick Karpathy-style if you want something running in 5 minutes.
loci/
README.md ← you are here
AGENT-SETUP.md ← agent onboarding protocol
PROCESSES.md ← agent-executable processes (garden-round, autodream, etc.)
FIRST-SESSION.md ← quickstart card
SETUP-GUIDE.md ← manual setup reference
templates/
CLAUDE-master.md ← master prompt template (with retrieval hierarchy)
SOUL.md ← soul file template
garden-template.md ← garden template (first-class)
persona-template.md ← persona creation template
scheduled-task-template.md ← morning briefs, garden rounds, comms digest
retrieval-hierarchy.md ← L0–L3 context loading protocol
room-template.md ← room context template
handover-template.md ← session delta format
tracker.json ← project tracker template
obsidian-mindmap-starter.md ← Obsidian canvas template
examples/
example-CLAUDE.md ← filled-in example
example-SOUL.md ← example soul file
modules/
[comms-integration]/ ← optional: drop a comms module here to pipe digests into morning check-in
The palace is plain text — markdown files and JSON. Any AI that can read files can run it. Switch engines between sessions, mix tools across accounts, use whichever model fits the task. The memory stays; the agent changes.
Tested and recommended:
- Claude — web, Cowork desktop, or API
- Claude Code — terminal (same palace, different surface)
- Claude across multiple accounts — work and personal share one palace seamlessly; only MCP integrations (Figma, Jira, etc.) differ by account
Works with any LLM that has file access:
- OpenAI GPT-4 / o3 — with file tools or Codex
- Google Gemini — via Drive or direct file context
- Mistral, Llama, Qwen — local via Ollama or any file-capable wrapper
- Any model via API with tool/function calling
The setup is the same regardless of engine. Point any AI at this repo, say "run the setup", and it interviews you and writes your palace. The CLAUDE-master.md template name is a convention, not a constraint — rename it if you're on a different stack.
session-deltaprocess — structured handover written at session close. Mandatory artifact listing (all files created/edited/deleted), TL;DR, state snapshot, decisions, open blockers, and exact next session opener. Established after a high-volume build sprint where implicit tracking was insufficient.- Website — loci.garden live. Public face of the methodology: palace map, three doors, dispatch archive, llms.txt agent declaration.
- Communication modules —
modules/zulip-crawler/generalised; docs now describe optional integrations for any team communication tool rather than Zulip-specific setup.
palace-updateprocess — delta analysis of user's palace vs. current Loci features. Verbose gap reports (why it matters, exact fix, effort estimate). Verbosity modes: full / quick / area-specific / summary.- Cherry-pick onboarding flow — Block 9 of
AGENT-SETUP.mdexpanded with four opt-in questions: morning check-in, autodream, skill eval cadence, insight decay rules. One question at a time.skipandskip allalways valid. Revisitable any time viapalace-update.
- ASCII logo — four rooms, one per letter. Letters drawn in the same box-drawing characters as the palace walls (
│ ┌─┐ └─┘ ───). One visual language throughout. - Engine-agnostic "Works with" — palace is plain text; any LLM with file access can run it. Works across multiple accounts (work + personal) seamlessly.
CLAUDE-master.mdis a naming convention, not a lock-in. - Changelog — added to README; covers v0.1 through v0.4.
- Naming ceremony — agent name moved to Block 8 (after garden + daily routine). Names are now shaped by what the agent has learned about the user, not offered cold at the start.
- Daily routine — new onboarding question asks how the user actually starts their day. Stored as a crystal; seeds every morning check-in with real context instead of a generic template.
- Autodream — weekly scheduled palace housekeeping (garden round + pattern scan + stale tracker check). On by default. Runs without you.
- Daily routine check-in — personalised morning brief process. Pulls from your comms tool and/or Jira if configured.
- Communication modules — optional integrations (Slack, Discord, etc.) for pulling digests into the morning check-in. Drop any compatible module into
modules/. - Cross-environment portability — palace is file-based; works identically across Claude Code, Cowork desktop, and web. Documented in onboarding and README.
- Garden first-class — garden moved from optional appendix to core feature. Competitive differentiator: no other co-intelligence scaffold has it.
- L0–L3 retrieval hierarchy — context loads in priority layers (soul identity → active state → room context → deep history). Documented in
templates/retrieval-hierarchy.md. - Persona templates — named thinking modes with their own soul files and gardens. Personas collaborate via tea sessions.
- Crystal expiry —
valid_until: YYYY-MM-DDfield added to crystal format. Prevents stale facts from calcifying as ground truth. - Scheduled tasks —
templates/scheduled-task-template.md: morning briefs, garden rounds, deep synthesis. Dynamic path finding (no hardcoded session paths). - Comparison table — honest positioning vs MemPalace (benchmarked, vector search) and Karpathy-style (simplest). Different tools for different needs.
- Renamed to Loci — was
palace-starter. Method of Loci. Classical, 4 letters.
- Initial release: room structure, context crystals, soul file, session deltas, CLAUDE-master template.
- 4-room default layout (Great Hall, Dev, Design, Hatchery).
- Basic handover format. Tracker JSON. Crystal tiers (◆ ◈ ◇).
Built by Hux × Vesper · 2026 · loci.garden "A collaborator, not a tool."