Alaz launches self-hosted AI memory

The clever bit here is not “memory” as a buzzword, it’s the discipline of turning noisy assistant sessions into structured, searchable knowledge that can actually survive across projects.

• –The 9-stage learning pipeline suggests Alaz is trying to curate memory quality, not just dump chunks into a vector DB.
• –Hybrid retrieval is the real moat: FTS, dense vectors, ColBERT, graph expansion, RAPTOR, and decay should outperform single-signal recall on messy coding history.
• –Self-hosting matters a lot for private codebases, but the operational footprint is still non-trivial once you add PostgreSQL, Qdrant, Ollama, and rerankers.
• –The Claude Code hook flow lowers adoption friction because memory only matters if it gets injected automatically at session boundaries.
• –The main risk is stale or overconfident context: persistent memory is useful only if contradictions and decay stay under control.