Homarus is an open research and engineering project for a poisoning-resistant, self-evolving long-term memory layer for personal and agentic AI systems.
The current design line is published as Lobster-Memory v0.5. It combines a fast wake path for interactive work, a slower local sleep path for memory consolidation, an ACE-style playbook for procedural experience, and a Memory Shield for provenance, quarantine, and adversarial regression testing.
This repository is documentation-first. The initial public release focuses on architecture, research references, contribution process, and cleaned whitepapers. Implementation code will be proposed through public issues and design discussions.
English is the canonical documentation language. Chinese editions are maintained as first-class companion documents for accessibility and community review.
- Keep personal memory inside a local trust boundary by default.
- Treat memory writes as security-sensitive events, not passive RAG ingestion.
- Preserve authoritative memories as natural language, with embeddings as disposable indexes.
- Make procedural experience auditable through structured playbook entries.
- Measure progress with memory, tool-use, and adversarial poisoning benchmarks.
Homarus is public and open to community help. Useful contributions include design critiques, reference corrections, benchmark reproductions, threat-model review, documentation improvements, and implementation proposals.
Please read CONTRIBUTING.md, governance, and SECURITY.md before opening large changes.
Homarus is licensed under the Apache License 2.0. Reference repositories are included as Git submodules and remain under their upstream licenses.