memory-palace

name	memory-palace
description	Manage durable cross-session memory via Memory Palace MCP server. Supports semantic search, write guards, and hierarchical memory organization. Use for saving facts, recalling context, and managing long-term knowledge. Chinese triggers: 记忆, 长期记忆, 记住, 回忆, 召回, 压缩上下文, 重建索引.

Memory Palace

Durable, semantically-searchable memory system for AI agents. Persists knowledge across sessions via a local MCP server.

Setup: See Memory Palace repository for installation. Requires Ollama with an embedding model (recommended: snowflake-arctic-embed2).

First Action in Any Session

Always start with:

read_memory("system://boot")

This loads core memories and recent context.

Core Workflow

1. Boot: read_memory("system://boot")
2. Search: search_memory(query, include_session=true) if URI is unknown
3. Read: read_memory(uri) to inspect the exact target
4. Mutate: create_memory or update_memory only after reading
5. Maintain: compact_context or rebuild_index when needed

Key Rules

• Read before write: Always read existing memory before creating/updating
• Respect guards: If guard_action=NOOP, stop and inspect the suggested target before proceeding
• Prefer update over create: Use update_memory when a related memory already exists
• Fresh context: In new sessions or subagents, always reload with system://boot

Available Tools

Tool	Purpose
read_memory	Read a memory by URI
search_memory	Semantic search across all memories
create_memory	Create a new memory
update_memory	Update an existing memory
delete_memory	Delete a memory
add_alias	Add an alternative URI for a memory
compact_context	Distill noisy session history
rebuild_index	Rebuild the semantic search index
index_status	Check index health

Common URIs

• system://boot — Session initialization
• system://index — Full memory index
• system://recent — Recently modified memories
• core://agent — Agent identity and rules
• core://my_user — User profile