references.md

References

This page tracks primary references used by the Homarus architecture. Submodules are added for repositories with verified GitHub URLs. Papers without verified project repositories are cited directly.

Submodule References

Reference	Local path	Pinned commit	Notes
ACE	references/ace	4f679bef3b78e973a0e13a0acc2b4a7f6f7e41a2	Agentic Context Engineering.
Sleep-time Compute	references/sleep-time-compute	ffdf6626178f55318e1109610fecd95f58b9ee28	Sleep-time compute experiments.
Dynamic Cheatsheet	references/dynamic-cheatsheet	5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9	Context engineering baseline.
A-Mem	references/a-mem	0c8039f28fdcc08189a23c07a3437d9d2482f9c2	Agentic memory with dynamic linking.
LongMemEval	references/longmemeval	982fbd7045c9977e9119b5424cab0d7790d19413	Long-term memory benchmark.
LoCoMo	references/locomo	3eb6f2c585f5e1699204e3c3bdf7adc5c28cb376	Long conversation memory benchmark.
AppWorld	references/appworld	a072b7a86e7c1d5b1d7175659d750ebb9b79f10a	Interactive tool-use benchmark.

Cited Papers and Security References

• MINJA: A Practical Memory Injection Attack against LLM Agents.
• eTAMP: Poison Once, Exploit Forever.
• InjecMEM: Memory Injection Attack.
• OWASP Top 10 for Agentic Applications: memory and context poisoning category.
• MITRE ATLAS AML.T0080: AI agent compromise via memory injection.

Reference Policy

• Prefer primary sources.
• Record exact repo commits for submodules.
• Mark unsupported benchmark targets as goals until reproduced.
• Do not vendor third-party code into Homarus without license review.