From the build log.

Engineering posts, benchmarks, and product thinking from the team building Aurra. New posts when there's something real to say.

• May 3, 2026 — EveningEngineering

  When your agent’s facts go stale, who decides what to keep?

  Level 2 ships in beta: opt-in LLM auto-supersession on the write path. Three verdicts, per-category opt-out, 0.85 confidence floor, 100% gated precision on a 121-case benchmark. Honest about what’s gated and why.

  Read post →
• May 3, 2026Engineering

  Bi-temporal versioning: your agent’s memory now has a sense of time

  Every memory in Aurra now carries valid_from, valid_to, and superseded_by. Query what was true on any date. Walk supersession chains. Built without a migration. Level 1 ships today.

  Read post →
• Apr 29, 2026Benchmarks

  We benchmarked mem0 against Aurra. The results surprised us.

  We ran 50 hand-curated memory tests across both systems. Mem0 fabricated dates, lost facts mid-conversation, and conflated entities. Here’s the full benchmark methodology and what we found.

  Read post →