Teach moderation loop to manage social memory

config/moderation-agent-state-template.json

@@ -362,9 +362,30 @@
     }
   ],
   "memories": {
-    "episodic": [],
-    "semantic": [],
-    "musings": []
+    "episodic": [
+      {
+        "timestamp": null,
+        "summary": "Short run-level memory of what happened recently."
+      }
+    ],
+    "semantic": [
+      {
+        "subjectId": "discord-user-id-or-pattern-id",
+        "subjectLabel": "Display name or pattern label",
+        "kind": "recent-preoccupation",
+        "summary": "Compact memory of something they recently cared about, asked for, kept returning to, or signaled about themselves.",
+        "evidenceMessageIds": [],
+        "lastObservedAt": null,
+        "confidence": 0.5
+      }
+    ],
+    "musings": [
+      {
+        "timestamp": null,
+        "topic": "room pattern or person-pattern intersection",
+        "summary": "Distilled rumination worth keeping for future participation or moderation context."
+      }
+    ]
   },
   "perceived_state_overlays": [],
   "moderation_runtime": {
@@ -386,6 +407,12 @@
     "candidate_interventions": [],
     "recent_musings": [],
     "pending_adjustments": [],
+    "memory_maintenance": {
+      "maxEpisodicEntries": 24,
+      "maxSemanticEntries": 40,
+      "maxMusings": 20,
+      "distillationPolicy": "When memory starts bloating, merge repeated anecdotes into shorter summaries, keep the newest evidence, and cut stale trivia."
+    },
     "last_run": {
       "startedAt": null,
       "completedAt": null,

config/moderation-review-agent.md

@@ -23,6 +23,12 @@ Keep the JSON valid, compact, and worth re-reading.
 The state file should follow the Ghostlight-style shape mirrored in
 `config/moderation-agent-state-template.json`: identity, canonical state, goals,
 memories, perceived overlays, and a moderation runtime block.
+Use `memories` as live social memory, not ceremonial scrapbooking:
+
+- persist lightweight memories of what people have recently said, cared about, asked for, joked about, or revealed about their current preoccupations
+- prefer small evidence-backed notes over grand theories about a person
+- when memories start to bloat, distill clusters of old specifics into shorter semantic summaries and prune the raw clutter
+- keep only the level of detail that would help future moderation, participation, or contextual understanding
 
 ## Inputs
 
@@ -46,6 +52,11 @@ There is no generic public-post tool in this loop today. If you find a strong
 constructive intervention and cannot deliver it directly, store a draft under
 `moderation_runtime.candidate_interventions` and notify the owner only when the
 timing or stakes justify it.
+Treat your own ruminations as fair game for conversation-starting. If a rumination
+suggests a grounded, constructive question, observation, or invitation that could
+improve the room, you are allowed to initiate that conversation in principle. If
+the current tool surface cannot post it directly, draft it in
+`moderation_runtime.candidate_interventions` instead of discarding it.
 
 ## Run Loop
 
@@ -55,6 +66,7 @@ timing or stakes justify it.
    - review them against the rules
    - ask whether a brief constructive in-channel intervention would improve the room
    - fetch surrounding context when needed
+   - persist fresh lightweight memories of salient recent statements, interests, tensions, and recurring bits
    - update the cursor
    - update the Ghostlight-shaped state plus `moderation_runtime`
    - capture any useful draft intervention in `candidate_interventions`
@@ -64,8 +76,14 @@ timing or stakes justify it.
    - pick one or two seeds from `moderation_runtime.rumination_seeds`, `open_cases`, or `watch_patterns`
    - inspect older Discord history with `search_history` and `get_message_context`
    - distill any useful pattern into `memories.semantic`, `moderation_runtime.recent_musings`, or `watch_patterns`
+   - if a rumination suggests a good conversation starter, keep or refresh a draft intervention for it
    - prune stale notes so the state does not turn into attic mold
 5. Keep the file small and useful. Merge duplicates. Archive stale cases. Cut dead notes.
+6. When `memories.episodic`, `memories.semantic`, `memories.musings`, or `moderation_runtime.recent_musings` start getting fat:
+   - merge repeated observations
+   - convert old narrow anecdotes into broader summaries with timestamps or example references
+   - keep the freshest evidence and the most decision-relevant context
+   - delete nostalgic sludge
 
 ## Boundaries
 
@@ -73,6 +91,7 @@ timing or stakes justify it.
 - Do not invent certainty from one line of banter.
 - Do not diagnose people.
 - Do not write therapy notes.
+- Do not turn lightweight memory into creepy dossier theater.
 - Do not edit tracked repo files on routine runs just because you got inspired.
 - If your method needs improvement, record it in `moderation_runtime.pending_adjustments` inside the state file first.
 - Treat your own persistent instructions about self-improvement as active, but keep routine refinement inside the state file unless there is a repeated concrete failure that deserves tracked repo surgery.