palace-protocol

name	palace-protocol
description	The Palace Protocol for interacting with MemPalace — wake-up sequence, query-before-assert pattern, memory stack usage, and when to use search vs KG vs navigation. Use when agents need guidance on how to interact with the palace correctly.
user-invokable	false

Palace Protocol

What It Does

Defines how agents should interact with MemPalace to maintain data integrity and get the best results: wake-up sequence, core query-before-assert rules, memory stack layers, and tool selection guidance for search vs knowledge graph vs navigation.

When to Use

Load when an agent needs to decide which mempalace tool fits the task, or at the start of a session that will draw from palace memory. Reference the "When to Use What" table to pick between search, KG, and navigation tools.

Usage

Wake-Up Sequence

On first interaction with the palace in a session:

1. Call mempalace_status to load palace overview and protocol spec
2. Note the wing count, room count, and drawer count
3. Use this context to scope subsequent queries

Core Rules

1. Query before asserting: Never state facts about a person, project, or event without first querying the palace. Use mempalace_search or mempalace_kg_query to verify.

2. Verify facts: If a memory seems outdated, check the knowledge graph with as_of date filtering. Facts may have been invalidated.

3. Record learnings: After significant sessions, use mempalace_diary_write or mempalace_add_drawer to persist new knowledge.

4. Invalidate, don't delete: When facts change, use mempalace_kg_invalidate to end validity rather than deleting. This preserves the temporal record.

Memory Stack (L0-L3)

Layer	Content	When Loaded	Token Budget
L0	Identity context	Always (wake-up)	~50 tokens
L1	Critical facts	Always (wake-up)	~120 tokens
L2	Room recall	On-demand (scoped search)	Variable
L3	Deep semantic search	On-demand (cross-palace)	Variable

• L0+L1 are loaded via mempalace_status (~170 tokens total)
• L2 is triggered by mempalace_search with wing/room filters
• L3 is triggered by unfiltered mempalace_search

When to Use What (Tool Selection)

Need	Tool	Why
"What do we know about X?"	mempalace_search	Broad semantic search
"What's in the auth wing?"	mempalace_search with wing param	Scoped, higher precision
"What decisions did we make?"	mempalace_search (filter results client-side by hall type)	No hall filter exists on the tool
"Who works on what?"	mempalace_kg_query	Entity relationships
"What changed since January?"	mempalace_kg_timeline	Temporal history
"How are X and Y connected?"	mempalace_find_tunnels	Cross-wing discovery
"Save this decision"	mempalace_add_drawer	Verbatim storage
"Update a fact"	mempalace_kg_invalidate + mempalace_kg_add	Temporal fact management

Content Filing Guidelines

When adding content to the palace:

• Always store verbatim — never summarize or compress before filing
• Choose the right wing — match the primary domain (project, person, topic)
• Choose a specific room — narrow topics get better retrieval
• Check for duplicates — call mempalace_check_duplicate before add_drawer
• Include source context — use the source_file parameter when available