Skip to content

Memory & self-learning

This is the feature that makes Prevail yours. A normal chatbot forgets you between sessions. Prevail remembers — and the more you use it, the more it knows your situation, so its answers get more specific over time.

you send a message
① intent captured raw transcript written immediately — never lost
② distilled the distillation daemon condenses chats into memory
③ _memory.md per-domain memory, plain markdown you own
④ read back fed into the next chat in that domain as context

The moment you hit send, Prevail writes the message — and the reply — to a durable intent ledger. This happens before anything else, so a raw transcript of what you asked is never lost, even if a model errors out mid-answer.

A background distillation daemon condenses each domain’s conversations into a compact summary: the facts, decisions, and context worth keeping. You don’t trigger this — it just happens.

The result is _memory.md inside each domain folder — plain markdown you own. Open it, edit it, correct it, or delete a line you don’t want remembered. It syncs with the rest of your vault.

On your next message in that domain, Prevail loads _memory.md as context. So when you ask a follow-up next week, it already knows your situation — you don’t re-explain.

Each domain shows how much context it’s carrying — recent chats plus distilled memory. It’s a quick read on how well Prevail “knows” that part of your life, and a nudge to prune or top up.

Context Score explains the underlying signal.

ArtifactLives inWhat it is
Intent ledger~/.prevail/ (machine-local)The raw, durable transcript of every turn
_memory.md<domain>/ (in your vault)Distilled per-domain memory, read back into chats
_log/<domain>/ (in your vault)A dated one-line decision log
_journal/<domain>/ (in your vault)Longer-form notes

All of it is local-first — nothing leaves your machine unless you send a message to a cloud model. Run a domain on a local model and the entire learning loop stays on-device.

Vault structure · Logs & journal · Privacy & local-first