| name | mempalace-architecture |
| description | Overview of mempalace architecture including wings, halls, rooms, BBBL, and palace structure |
Palace structure
A mempalace instance contains all stored observations organized in a hierarchy:
/path/to/palace/
mpr_drawers.json -- main embedvec collection (observations)
mpr_compressed.json -- compressed versions (AAAK dialect)
tunnels.json -- mesh peer connections
Wings, halls, and rooms
Data is organized by topic (wing) and sub-topic (hall/room), following the memory palace metaphor:
- Wing -- broad topic area (e.g., "emotions", "technical", "identity"). Detected from directory names during
mpr init.
- Hall -- category within a wing ("bug-fixes", "decisions", "patterns").
- Room -- specific context or file within a hall.
Default wings: emotions, consciousness, memory, technical, identity, family, creative.
BBBL (Basic-Building-Block-Language)
AAAK Dialect for compressing observations to ~3-5% of original size:
@KAI @PRI
&project (mempalace_rust | config)
=KAI We decided to use XDG_CONFIG_HOME ... (config-resolution | 2026-06-10)
This is an identity-aware shorthand where people are mapped to codes (KAI, PRI) and observations are compressed with type markers (@, &, =).
Memory layers
- L0 Identity (
identity.txt): AI's self-concept, loaded every session.
- L1 Wake-up: Recent context from the last few sessions (~600-900 tokens).
- L2 Recall: On-demand search via hybrid BM25 + vector + graph.
- L3 Consolidation: Periodic LLM-based compression of observations into memories.
- Graph: Knowledge graph with entities, relations, and graph-expanded search.
Storage
- Config:
~/.mempalace/config.json (XDG-compliant, $XDG_CONFIG_HOME/mempalace/)
- Palace data: Default
~/.mempalace/palace/ (configurable via palace_path)
- Identity:
~/.mempalace/identity.txt
- Entity registry:
~/.mempalace/entity_registry.json
- People map:
~/.mempalace/people_map.json