v0.21.0

Memory is now something you distill on purpose. Nothing reads a session transcript on its own any more — /iwe:distill walks you through what a session established and writes only what you choose to keep, and session start tells you how many sessions are still waiting. Documents written into the store are normalized and checked against their schemas as they land, and iwe normalize -k <key> cleans up just the pages you name.

iwe

Added

• iwe normalize -k <key> normalizes only the named documents, leaves their frontmatter as written and prints only the paths that changed
• knowledge_filter, recency_field, injection, max_proposals_per_read and remind_every_days knobs on MEMORY.md
• Switching memory on installs a default .iwe/schemas/memory.yaml that constrains only the created and session fields
• A document written with the Write or Edit tool is normalized, checked against the schemas that bind it, and reported when it closely matches one the store already has
• A captured document under an area directory is linked into that area's hub when one exists
• Session start reports how many sessions are undistilled and, at most once per remind_every_days, offers to read them

Changed

• update block operators (--append, --replace, --delete, …) given an argument that is not a YAML mapping now report the flag, the mapping shape it takes and an example (previously the raw parser error)
• Memory's own state lives under .iwe/claude/, outside the graph (previously sessions/<id> documents in the store). The default knowledge_filter is { $key: { $nin: [MEMORY, queries] } } (was { distilled_lines: { $exists: false }, $key: { $nin: [MEMORY, queries] } })
• The init, distill and reflect prompts carry procedure only; the shape of a memory document is the MEMORY.md policy's alone, read by section name (## How to write it, ## Dedup and updates, ## Curation)
• /iwe:distill asks about every candidate before anything is written — Keep all / Skip all / Let me pick, then one candidate at a time for a handful, or a numbered list with batched multi-select questions for more — and reads a backlog with one reader per session, so a fact several sessions raised is asked about once (previously it walked the backlog session by session and re-proposed what recurred)
• Reading a session can now run unattended; selection never does
• /iwe:init sets memory up and hands the sessions on disk to /iwe:distill; it reads none of them itself
• The session-start block's closing lines come from the policy's ## At session start section (previously hard-coded)
• iwe create and iwe update --content normalize the body on the way in, leaving the frontmatter as written

Removed

• Automatic capture: nothing reads a transcript on its own any more, and the chunks it kept under .iwe/claude-sessions/ are gone
• The sweep_threshold_lines, max_chunks_per_sweep, max_items_per_chunk and inflight_ttl_minutes knobs; a policy that still sets one is ignored

Library crates (liwe, diwe)

liwe

Added

• parse_filter_mapping — build a query Filter from an already-parsed YAML Mapping

Changed

• $set accepts reserved-prefix keys ($, _, ., #, @) inside a value; the field path itself is still refused when a segment starts with one