MemPalace

This note profiles MemPalace, a local-only memory system that stores raw exchanges instead of summaries. It starts with the claim and the headline numbers, then gives the three benchmark figures, including the LongMemEval R@5 result and the zero annual cost. The feature list explains the palace metaphor, the verbatim storage choice, the experimental abbreviation dialect and the SQLite knowledge graph. Install and CLI commands come next, followed by a short Python call and the MCP server entry point. Later sections cover the architecture, what each storage layer holds, the use cases, and why the tool stays free. Pros and cons close the page, along with the situations where MemPalace is the right pick.

Definition

MemPalace is advertised as the "highest-scoring AI memory system ever benchmarked. And it's free." With 44k stars, it is one of the strongest memory systems for AI assistants.

Key Metrics

• Stars: 44k
• Forks: 5.6k
• License: MIT

Benchmarks

Key Features

Palace Structure

• Wings: People/Projects
• Rooms: Topics
• Halls: Memory types
• Tunnels: Cross-wing connections
• Closets: Summaries
• Drawers: Original files

Raw Verbatim Storage

• Stores the actual exchanges in ChromaDB
• No summarization
• Achieves 96.6% LongMemEval R@5

AAAK Dialect

• Experimental lossy abbreviation system
• Token compression
• Currently regresses to 84.2% vs raw mode

Knowledge Graph

• Temporal entity-relationship triples in SQLite
• Local alternative to Zep's Graphiti

Specialist Agents

• Create focused agents with their own wings
• Individual diaries

MCP Integration

• 19 tools available
• Claude, ChatGPT, Cursor, Gemini support

Installation

pip install mempalace
mempalace init ~/projects/myapp
mempalace mine ~/projects/myapp

CLI Commands

# Initialize
mempalace init ~/projects/myapp

# Mine data
mempalace mine ~/projects/myapp

# Search
mempalace search "query"

# Status
mempalace status

# Wake-up
mempalace wake-up

# Split
mempalace split

Python API

from mempalace.searcher import search_memories

# Search memories
results = search_memories(
    query="authentication bug",
    limit=10
)

MCP Server

python -m mempalace.mcp_server

Architecture

┌─────────────────────────────────────────┐
│           MemPalace                     │
├─────────────────────────────────────────┤
│  ┌─────────────┐  ┌──────────────────┐  │
│  │  ChromaDB   │  │     SQLite      │  │
│  │  (Vectors)  │  │ (Knowledge Graph)│ │
│  └─────────────┘  └──────────────────┘  │
├─────────────────────────────────────────┤
│           MCP Server                    │
│  (19 tools for AI integration)          │
└─────────────────────────────────────────┘

Storage Details

Use Cases

• Solo developers: Managing multiple AI conversations across projects
• Team leads: Tracking decisions, preferences, project history
• Anyone wanting persistent AI memory: Without cloud dependency

Why It's Free

• Local-only (no cloud, no API calls)
• Zero external dependencies
• Self-hosted

Tech Stack

• Python 3.9+
• ChromaDB (vector storage)
• SQLite (knowledge graph)
• MCP for AI integration
• Local-only (no cloud, no API calls)

Pros

Cons

When to Use

• Maximum accuracy: When benchmark performance is the priority
• Privacy-first: You do not want a cloud dependency
• Cost-sensitive: Free with the full feature set
• Multi-project: Managing multiple AI conversations
• Team memory: Shared knowledge across a team

MemPalace stores the raw exchange in ChromaDB and skips summaries, which is the mode behind the 96.6% LongMemEval R@5 score on all 500 questions. The palace layout splits people and projects into wings, topics into rooms, and memory types into halls. A SQLite graph keeps temporal links between entities, as a local alternative to Zep's Graphiti. The experimental abbreviation dialect falls back to 84.2%, so raw mode stays the accurate path.

References:

• MemPalace/mempalace
• Documentation
• Benchmarks

See also: Memory Tool Comparison · Mem0