zer0dex is a local two-tier memory pattern reference implementation for AI agents, published by hermes-labs-ai on GitHub. Its core goal is to give agents long-term recall capability without loading all notes into the prompt.
Purpose and positioning: when an agent knows the project name, but the deployment decision it actually needs is buried in notes accumulated over months, putting the entire archive into context is both expensive and hard to retrieve; and using only a vector store lacks a simple map that lets you directly see what is in it. zer0dex pairs a short, readable and editable index with semantic search over the underlying memory, solving this cross-project recall problem. According to the official description, it is suitable for developers building local, long-running agents whose notes have already exceeded what a single MEMORY.md can hold, but who still want to keep a readable map of what has been saved. It is a reference implementation of the two-tier pattern, currently on the 0.1.x developer preview line; it is not a hosted team service, nor a complete agent framework.
Main capabilities and two-tier structure: the first tier is a compact, human-readable Markdown index used to show what the agent knows; the second tier is a local vector store that retrieves specific content when details matter. The command-line tool can seed memory from Markdown, append more text, and query a loopback HTTP server. The default path uses local Ollama models for embedding and memory extraction, with mem0 and Chroma as storage. Configuration and storage are located in the current directory (.zer0dex.json and .zer0dex/). Users can inspect the index itself and keep the retrieval store on their own machine.
How to use: before running, Python 3.11 or 3.12 is required, and local Ollama must already be http://localhost:11434 Provide the service; if it is not yet running, Ollama must be started first. The official materials mention installation via PyPI and provide a CLI reference and HTTP API documentation. The agent host calls POST /query before initiating a model request and decides for itself which matching results to include in the context. zer0dex only provides the server and CLI and does not automatically install hooks in your agent; the repository includes a small TypeScript host adapter to demonstrate this integration point. The command examples in the materials show that the last command returns the matched stored memories when extraction and retrieval succeed; after completion, run zer0dex stop. During trial runs, it is recommended to use a disposable directory because the configuration and storage are written to the current directory. If the check fails, confirm that Ollama is running and that both models have been pulled before seeding.
Use cases and evaluation: The project is positioned for cross-project recall and is suitable for situations where flat memory files or pure vector RAG are not sufficient. The official materials provide a bundled evaluation: 86 memories, 97 use cases, and the combined method achieved 91.2% average recall and 80.0% cross-reference recall. The official materials also state that this is only a single test workload and that you should validate it with your own notes and queries before choosing a retrieval strategy.
Limitations and licensing: zer0dex is a reference implementation, belongs to the 0.1.x developer preview version, and is not a hosted team service, nor is it a complete agent framework. It requires Ollama to run locally, depends on Python 3.11 or 3.12, and by default uses the local Ollama model as well as mem0 and Chroma storage. Regarding charges, pricing, specific platform support scope, additional performance metrics, or more detailed usage steps, the official materials do not currently provide them. The project uses the Apache-2.0 license and is built by Hermes Labs as agent infrastructure for autonomous systems.
