agent-working-memory 0.16.1 → 0.18.1

package/plugin/skills/awm-memory/SKILL.md

@@ -10,7 +10,7 @@ description: >-
 
 # Agent Working Memory
 
-Generated from `src/adapters/common.ts` (AWM v0.16.1) — the same guidance
+Generated from `src/adapters/common.ts` (AWM v0.18.1) — the same guidance
 `awm setup` writes into CLAUDE.md. Do not edit here; edit the source and re-run
 `npm run build:plugin`.
 
@@ -157,7 +157,8 @@ people/things you refer to by name often); not yet the default pending evaluatio
 `memory_write` accepts `valid_from` / `valid_to` (ISO dates). Use `valid_to` on
 **operational** facts with a real shelf life — a deploy state, "waiting on X's reply",
 a ticket status — so the memory expires instead of relying on you to remember it's
-stale. Recall renders `[valid until …]` on results carrying this field. Use
+stale. Recall renders `[valid until …]` while the window holds; once it has passed
+the memory is down-ranked like a superseded one and labelled `⚠ EXPIRED`. Use
 `valid_from` for a fact that becomes true on a known future date (a policy change, a
 season that hasn't started yet). Don't set either for durable facts — most memories
 don't need them.
@@ -174,9 +175,11 @@ don't need them.
 ### Salience auto-promotion (defense in depth)
 The salience filter automatically promotes certain content patterns even if you forget
 to set `memory_class` explicitly:
-- **User feedback** — content starting with "Robert said…", "Katherine directed…",
-  "Nancy decided…" etc. auto-promotes to canonical. So quoting the user verbatim
-  always preserves the decision.
+- **User feedback** — content starting with a configured stakeholder name plus a feedback
+  verb ("<name> said…", "<name> directed…", "<name> decided…") auto-promotes to canonical,
+  so quoting the user verbatim preserves the decision. **This is OFF unless an operator
+  sets `AWM_FEEDBACK_NAMES`** (it ships empty); when in doubt, pass
+  `memory_class: 'canonical'` explicitly rather than relying on it.
 - **Verified operational records** — content with an action verb (Submitted, Finalized,
   Completed, Reconciled, Triaged, Posted, Resolved, Stamped, Pushed, Deployed, Migrated,
   Imported, Exported, Backfilled) plus 2+ concrete identifiers (ISO date `YYYY-MM-DD`,
@@ -283,6 +286,14 @@ the cost/quality tradeoff. Use them when the default doesn't match what you need
 ### Keep memory fresh
 - After recalling a memory, if you observe the real state is different → call
   `memory_supersede` immediately with the corrected version.
+- When a fact has CHANGED (a time, a price, a status, a decision) and you don't have the
+  old memory's id: `memory_write` with `correction: true`. AWM supersedes the closest
+  memory about the same thing and tells you which. Written as a plain observation, a
+  changed value merges into or ranks below the stale version — that is how "departs 8:40"
+  lost to "departs 10:15" in the MWA assessment.
+- A recall line that says `⚠ POSSIBLE CONFLICT`, `⚠ EXPIRED` or `⚠ SUPERSEDED` is
+  telling you not to state that value from memory: read the source, then supersede or
+  correct the stale one.
 - After using a recalled memory: call `memory_feedback` (useful/not-useful) **with the
   `recall_id`** printed at the end of the recall output (`[recall_id: …]`). That id
   joins the feedback to the recall that produced it, which is what lets AWM learn