The layer that makes everything else possible.
Hydrate is a local-first memory layer your agents share. Every fact carries provenance, who or what wrote it, when, and how often it has been reinforced, and recall now ranks by trust × freshness × relevance rather than raw similarity. Memory is a layer, not a feature: separable, queryable, editable and portable across models and tools.
Provenance and trust-weighted recall.
Recall ranks by trust × freshness × relevance, so well-sourced, repeatedly-confirmed decisions surface above one-off noise. This is the first memory-evolution default to flip on, after a calibration pass and a fix for an early ranking-collapse regression. The store owns the data model, provenance and capture-time claim management; everything runs locally, standalone.
opt-in preview Coming, behind flags
Three further units are shipped but default-off while the cross-runtime benchmark calibrates. Safe to try locally; not yet a promise, and not on by default.
Capture-time claim management
At capture time, decide Add / No-op / Update / promote-to-canon against what is already believed, so memory consolidates instead of accreting duplicates.
HYDRATE_AUDN Synthesised user model
A regenerated "who is this user" model, role, preferences, working style, served to the identity surface. Degrades cleanly to the legacy identity block when off.
HYDRATE_USER_MODEL Skills memory
Learn repeated multi-step procedures, store them with per-version trust, and inject the relevant skill when a matching task appears, bounded so skills never displace canon.
HYDRATE_SKILLS Local-first. No secrets in the chat.
The store lives on your machine. The daemon now auto-launches at login, so the memory layer is up before your first prompt. Nothing is sent to a cloud Hydrate operates for recall to work.