Homarus is an open-source effort to design and build a trustworthy long-term memory layer for AI agents. The project starts from the Lobster-Memory architecture line and publishes a cleaned, community-oriented v0.5 design.
Current agent memory systems often treat memory as passive retrieval augmented generation: collect text, embed it, retrieve similar chunks, and inject them back into context. That approach is useful, but it leaves important gaps:
- memory writes can be poisoned by direct or indirect prompt injection;
- procedural experience is hard to audit or remove once it is embedded;
- model upgrades can invalidate embedding-dependent memory formats;
- long-context models reduce but do not remove the need for durable, inspectable memory;
- benchmark claims are often not connected to reproducible regression tests.
Homarus explores four design commitments:
- A local trust boundary for raw personal data and consolidation work.
- A wake path for low-latency interaction and a sleep path for slower memory digestion.
- A Memory Shield that treats every memory write as a security event.
- An ACE-style playbook where procedural knowledge is structured, counted, sourced, and removable.
This repository is documentation-first. The first public milestone is a clear architecture, a shared vocabulary, vetted references, and open discussion. Code will follow after the public interfaces, threat model, and evaluation harness are agreed.
- Agent and memory-system engineers can review architecture feasibility.
- Security researchers can stress-test the poisoning and provenance model.
- Benchmark contributors can reproduce LongMemEval, LoCoMo, AppWorld, and injection test suites.
- Bilingual contributors can improve English and Chinese documentation quality.