Apis adds self-updating memory, unblocks LoRA training
Apis is a fully offline AI system running on Ollama and written in Rust. According to the maker, it expanded its Turing Grid memory from a single metadata cell to three populated cells, fixed a race condition in the training pipeline with semaphore locks, improved batch ordering, and successfully trained its first consolidated memory adapter overnight. The post also says the system reviewed its Voice subsystem, Kokoro TTS integration, and NeuroLease mesh discovery code, then kept running after recompilation without manual intervention.
Bold claim, but the practical takeaway is straightforward: this is a local-first agent stack with persistent memory and self-maintenance loops, not just a chat wrapper.
- –The update centers on reliability work: race-condition fixes, semaphore locking, and batch-order optimization are the kind of changes that make autonomous runs more credible.
- –The memory architecture expansion suggests the project is treating state as a first-class subsystem rather than ephemeral session context.
- –Open-source Rust plus Ollama makes the pitch attractive for offline and privacy-sensitive setups.
- –The self-modification angle is interesting, but the evidence here is a maker post, so I’d treat the autonomy claims as reported behavior rather than independently verified capability.
DISCOVERED
12d ago
2026-03-31
PUBLISHED
12d ago
2026-03-31
RELEVANCE
AUTHOR
Leather_Area_2301