Mnemosyne
Provides persistent, graph-based memory for AI agents via MCP, enabling semantic search, wikilink traversal, reminders, and injection protection.
README
🧠 Palimpsest
The Context Engine for AI Agents — Give your AI agents persistent, observable, compliant memory.
All memories are plain markdown files you own. Search by meaning, traverse relationships, schedule reminders, and protect against poisoned data.
⚠️ Rebrand Notice
This project was formerly known as Mnemosyne. We rebranded to Palimpsest in July 2026 to avoid confusion with an unrelated project that adopted the same name.
Why Palimpsest? A palimpsest is a manuscript on which later writing has been superimposed on earlier writing — yet traces of the original remain. It's the perfect metaphor for layered, persistent, evolving memory.
What is Palimpsest?
Palimpsest is a production-grade memory platform for AI agents. Unlike simple chat history or RAG, it gives agents:
- Long-term persistent memory — survives restarts, works across sessions
- Semantic search — find ideas by meaning, not just keywords
- Graph memory — notes link via
[[wiki-links]], traverse relationships - Security gates — injection detection, contradiction flagging, near-duplicate checks
- Prospective memory — "remind me in 3 days" — and it actually happens
- Sleep consolidation — nightly maintenance: archive stale, merge duplicates
- MCP server — Claude Code, Cursor, any MCP client can read/write memory
- Observability — audit trails, contamination detection, memory health dashboard
- Compliance — GDPR Article 17, EU AI Act ready
Architecture
┌─────────────────────────────────────────────┐
│ Your Question │
│ "What did we decide about API rate limit?" │
└──────────────────┬──────────────────────────┘
│
┌──────────────────▼──────────────────────────┐
│ Palimpsest Memory Platform │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Semantic │ │ Keyword │ │ Graph │ │
│ │ Search │ │ Search │ │ Search │ │
│ │(pgvector)│ │(tsvector)│ │(wikilinks│ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ └────────────┬────────────┘ │
│ │ │
│ ┌───────▼────────┐ │
│ │ RRF Merge │ │
│ └───────┬────────┘ │
│ │ │
│ ┌─────────────────▼────────────────────┐ │
│ │ Markdown Vault (source of truth) │ │
│ │ ~/Palimpsest/vault/*.md │ │
│ └─────────────────────────────────────┘ │
└─────────────────────────────────────────────┘
Quick Start
# Clone the platform
git clone https://github.com/M4F-S/palimpsest
cd palimpsest
# Install (SQLite works out of the box)
pip install -e ".[dev]"
# Or with PostgreSQL + pgvector for production:
docker run -d --name palimpsest-pg -p 15432:5432 \
-e POSTGRES_USER=palimpsest \
-e POSTGRES_PASSWORD=palimpsest_secret \
-e POSTGRES_DB=palimpsest \
ankane/pgvector:latest
from palimpsest import UnifiedMemorySystem
memory = UnifiedMemorySystem()
# Save a memory
memory.remember(
title="API Rate Limit Decision",
content="100 req/min with burst to 200. Alert if p95 > 200ms.",
tags=["api", "decision"],
salience=0.9
)
# Search by meaning
results = memory.recall("rate limiting policy", mode="hybrid", top_k=5)
# Schedule a reminder
memory.remind_me("Review API metrics", "2026-07-07T09:00:00", recurring="weekly")
# Run nightly maintenance
memory.consolidate()
Kimi Skill
For the Kimi AI skill (minimal installable version):
👉 github.com/M4F-S/kimi-palimpsest-skill
The skill repo contains only the essential files for Kimi integration: SKILL.md, palimpsest/ package, and tests.
Documentation
| Document | Purpose |
|---|---|
| PLATFORM_BUILDER_BRIEF.md | Architecture and design decisions |
| SETUP.md | Installation and configuration guide |
| PALIMPSEST_V3_REBUILD_PLAN.md | V3 roadmap and rebuild strategy |
| PALIMPSEST_STRATEGIC_DECISIONS.md | Strategic decisions log |
| RESEARCH_*.md | Research reports and competitive analysis |
| TEST_REPORT.md | Test coverage and results |
| CHANGELOG.md | Version history |
| CONTRIBUTING.md | Contribution guidelines |
Environment Variables
| Variable | Default | Purpose |
|---|---|---|
MEMORY_DB_DSN |
(none) | PostgreSQL connection string |
MEMORY_SQLITE_PATH |
~/.palimpsest/palimpsest.db |
SQLite database path |
MEMORY_VAULT_PATH |
~/Documents/Kimi/Workspaces/Palimpsest/vault |
Markdown vault directory |
EMBEDDING_MODEL |
all-MiniLM-L6-v2 |
Sentence-transformers model |
OLLAMA_URL |
http://localhost:11434 |
Ollama server for embeddings |
License
Apache 2.0
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。