为人工智能代理提供持久语义记忆。存储对话中的事实、决策、洞察和上下文。按意义搜索，而不仅仅是关键词。构建连接记忆的知识图谱。跨模型、实例和提供者共享。

问题

每次AI会话都从一张白纸开始。上下文窗口是有限的。会话结束，知识消失。

目前的解决方案都是厂商锁定的：ChatGPT 的内存只支持 OpenAI。Anthropic 的项目只能支持 Claude。换供应商你就得重新开始。你积累的上下文——决策、偏好、项目历史——属于供应商，而不是你。

与此同时，行业争相构建更大的上下文窗口。12.8万。20万。100万个代币。但你不能通过给某人更大的白板来解决人类失忆症。内存不属于模型内部——它应该与模型并存。

记忆宫殿是一个持久的语义记忆层，任何兼容MCP的人工智能都能访问。它将内存与模型分离，就像几十年前数据库将数据与应用分离一样。上下文窗口变成了工作内存。记忆宫殿是长期存储。这就是真正的大脑的工作原理。

快速启动

Memory Palace

Persistent semantic memory for AI agents. Store facts, decisions, insights, and context across conversations. Search by meaning, not just keywords. Build a knowledge graph of connected memories. Share across models, instances, and providers.

The Problem

Every AI session starts as a blank slate. Context windows are finite. Sessions end, knowledge dies.

Current solutions are all vendor-locked: ChatGPT’s memory only works with OpenAI. Anthropic’s projects only work with Claude. Switch providers and you start over. Your accumulated context — decisions, preferences, project history — belongs to the vendor, not to you.

Meanwhile, the industry races to build bigger context windows. 128K. 200K. 1M tokens. But you don’t solve human amnesia by giving someone a bigger whiteboard. Memory doesn’t belong inside the model — it belongs alongside it.

Memory Palace is a persistent semantic memory layer that any MCP-compatible AI can access. It separates memory from the model, the same way databases separated data from applications decades ago. The context window becomes working memory. Memory Palace is long-term storage. That’s how actual brains work.

Quick Start

# Clone and install
git clone https://github.com/jeffpierce/memory-palace.git
cd memory-palace
pip install -e .

# Run setup wizard (detects GPU, downloads models)
python -m setup.first_run

Platform-specific installers are also available: install.bat / install.ps1 (Windows), ./install.sh (macOS/Linux).

See docs/README.md for detailed installation and configuration instructions.

Requirements

• Python 3.10+
• Ollama (local model serving)
• NVIDIA GPU with 4GB+ VRAM (recommended, not required)

Model Selection

Models are auto-detected by the setup wizard. Defaults are chosen to run everywhere:

See docs/models.md for the full model guide with VRAM budgets and upgrade paths.

Features

• Semantic Search — Find memories by meaning using local embedding models via Ollama
• Knowledge Graph — Typed, weighted, directional edges between memories with automatic graph context in retrieval
• Centrality-Weighted Ranking — Retrieval scores combine semantic similarity, access frequency, and graph centrality
• Auto-Linking — New memories automatically link to similar existing ones (configurable thresholds)
• Multi-Project Support — Memories can belong to multiple projects simultaneously