Dave

We spent over $100K (compute, dead ends, the lot) and trained on 49M examples across about 3M synthetic life scripts to get a 4B memory specialist that runs fully local. Extract, temporal, reconcile, abstain. LoRA on a 4B base, INT4 is a couple GB on disk, a few GB hot. Comfortable on a 16GB laptop. Phone packaging soon and what we're working on next. This isn’t just “smarter chat.” I see it as a deployment bet: memory as a software component on the machine that holds the data. No per-write API tax. Nothing leaves the device. Frontier labs memory remembers you inside their product. Coding agents already get rich context. Your health notes, dose changes, shift logs, ambient scraps with other people’s names in them? That stuff either lives in a silo or nowhere. We want the model next to the notes you own, something that is really personalised. Security context is the quiet part. CONTINUITY’s point is that agent stacks drop, widen, or reinterpret authorization as work crosses tools. Private memory only works if provenance survives the whole path: who wrote it, what it authorizes, what you’re allowed to do when two records disagree. Reconcile and abstain aren’t features, it's integrity and something we've pushed from a Private AI perspective. Imagine; 1️⃣ Offline plant handover where two valve readings conflict and there is no outbound net. Surface both. Don’t silently keep the newer row. 2️⃣A clinician’s notes on a device that can’t phone home. Dose flipped twice, third source disagrees. Refuse the confident answer. 3️⃣Ambient capture on a laptop. Thousands of fragments a day. High write volume plus privacy is where local stops being a hobby and starts being the only sane architecture. We also killed our own vanity number. Hit 100% on our bench. Four regexes hit 100% too on 2.7M training examples. We rebuilt so phrasing carries no label. Bench ships with an attack: bag-of-words above chance means we automatically reject it. Model lands at 41.4%. Shortcuts around 25%. Weights + gated bench going up on Hugging Face soon. Come try to break it.

Clude trained the world’s first memory model that remembers better than frontier models 100x its size. Read more here:

Clude Memory Model Update: We’ve now trained our 4B specialist memory model on 4.9M examples across nine memory operations. The results are encouraging. Clude scores 89.7 across our MemOpsEval benchmark, ahead of Claude Sonnet at 88.9 and Gemini Flash at 85.0. More importantly, it performs best where persistent memory tends to break; temporal reasoning and reconciling contradictory information. We’ve also learned a few things along the way. Performance largely saturated around 1.2M training examples. Splitting evaluation by synthetic life history, rather than individual examples, matters enormously. The model now runs quantized locally, with no per-memory-operation API cost. Next is per-user adapters and continuous refresh from confirmed memories and corrections. We’ll publish a more detailed technical write-up update after.