Description
CheeseDB is a specialized database engine engineered for AI memory management, emphasizing source traceability and selective forgetting. Unlike traditional databases, CheeseDB ensures that every piece of information, or 'memory,' is linked directly to its original source. This fundamental design addresses a critical problem in AI agent development: when an untrustworthy source is identified, simply deleting the source isn't enough; all memories derived from it must also be purged to prevent the agent from continuing to act on incorrect, malicious, or private data.
Read More
The engine allows users and AI agents to interact with memories using plain English queries. Leveraging an embedding model, CheeseDB finds memories by meaning, eliminating the need for exact keyword matches. When a memory is recalled, it is permanently removed from search results, ensuring that agents only access valid information. The system supports various OpenAI-compatible embedding servers, including local options like Ollama, LM Studio, oMLX, and vLLM, or hosted APIs, allowing users to keep their memories on their own machines for enhanced privacy.
Under the hood, CheeseDB structures each record into three parts: a header for ID and hash-chaining, a structure for type and source dependencies, and the content itself. This architecture ensures that even after content is erased, the history and dependencies remain verifiable, preserving the integrity of the memory chain. The content, along with its search vector, is the only part removed during an erase operation, making the deletion truly permanent from the files.
For developers, CheeseDB offers a Go library for direct integration and a local MCP server for applications like Claude Code and Cursor. It provides tools for creating indexes, embedding data, and searching. The command-line tool allows for easy installation and interaction, including a `find` command to verify that erased text is truly gone. While currently in preview, CheeseDB is ideal for scenarios where AI agents learn from untrustworthy sources, require deep chain recall, need plain English search, demand verifiable data erasure, or integrate with MCP apps without extensive coding. It is developed on macOS and tested on Linux, with Windows support pending. The name 'CheeseDB' is inspired by food safety practices, where contaminated lots of milk lead to the recall of specific cheese wheels, mirroring how the database recalls memories linked to bad sources.
- CategoryVector Databases & Retrieval
- DeploymentEmbedded
- PlatformsmacOS, Linux
CheeseDB's Core Features
Source-aware memory management for AI agents
Plain English querying for memories
Selective forgetting by describing what to remove
Hash-chained memory database for data integrity
Embedded Go library for direct integration
Local MCP server for Claude Code, Cursor, and other MCP apps
Support for OpenAI-compatible embedding servers (Ollama, LM Studio, oMLX, vLLM)
Memory search by meaning, not just keywords
Verifiable erasure of content while preserving history
Append-only segments for agent data on disk
Rebuilds family tree, leg counters, and offset table from log
Command-line tool for interaction and verification
Developed on macOS and tested on Linux
No external model or network required for core functionality
Integration with existing vector stores by storing memory IDs
How to use CheeseDB?
Install CheeseDB: Use `go install github.com/govinargo-labs/cheesedb/cmd/cheesedb@latest` to get the command-line tool.
Configure Embedding Model: Add two lines to `~/.cheesedb/llm.conf` to specify your OpenAI-compatible embedding server (e.g., Ollama, LM Studio).
Remember Memories: New memories become searchable as they are remembered; use `cheesedb embed alice` to catch up on older ones.
Search Memories: Ask questions in plain English using `cheesedb search [agent_name] "your question?"` to find relevant memories.
Get Answers: Use `cheesedb ask [agent_name] "your question?"` to have a chat model answer from the closest live memories.
Forget Memories: Describe what to forget with `cheesedb forget [agent_name] "description"`; confirm erasure after review.
Verify Erasure: Use `cheesedb find` to search raw bytes and confirm that erased text is truly gone from the files.
CheeseDB's Use Cases
- Managing Untrustworthy AI Sources
- Deep Chain Memory Recall
- Plain English Memory Search
- Verifiable Data Erasure
- Integrating with MCP Apps
- Embedded Go Library Solutions
- Preventing Memory Poisoning
- Handling User Data Deletion Requests





