holoduke · holoduke · Jun 1, 2026 · Jun 2, 2026
diff --git a/backend/brain-prompt.ts b/backend/brain-prompt.ts
@@ -2,6 +2,7 @@ import type { Observation } from "./observer.js";
 import type { MemoryNode, WorkingMemory } from "./memory/types.js";
 import type { MemoryGraph } from "./memory/graph.js";
 import { serializeNodesForPrompt, collectRelevantRejectedEdges, formatRejectedEdgesForPrompt } from "./memory/activation.js";
+import { trackingItemText } from "./memory/working-memory.js";
 import { ariaPersonality } from "./aria-identity.js";
 import type { CharacterOverride } from "./aria-identity.js";
 import { getBrainConfig, getCharacterPreset, getOwnerLocalTime } from "./brain-config.js";
@@ -287,7 +288,7 @@ function formatWorkingMemory(wm: WorkingMemory): string {
   const parts: string[] = [];
   if (wm.currentContext) parts.push(`Context: ${wm.currentContext}`);
   if (wm.mood) parts.push(`Mood: ${wm.mood}`);
-  if (wm.shortTermTracking?.length > 0) parts.push(`Tracking: ${wm.shortTermTracking.join(", ")}`);
+  if (wm.shortTermTracking?.length > 0) parts.push(`Tracking: ${wm.shortTermTracking.map(trackingItemText).join(", ")}`);
 
   // Temporal context
   if (wm.temporal) {
@@ -708,7 +709,7 @@ Respond with ONLY a JSON object:
   "workingMemory": {
     "currentContext": "brief summary of what's happening right now",
     "mood": "your current mood/energy",
-    "shortTermTracking": ["things you're actively watching"],
+    "shortTermTracking": ["things you're actively watching — see EVIDENCE-LINKING below for resolved items"],
     "pendingFollowUps": [{"id": "fu_8hex", "question": "what to follow up on", "targetPerson": "name or omit", "context": "why", "createdAt": ${Date.now()}, "dueAt": null}],
     "conversationThreads": []
   },
@@ -755,6 +756,9 @@ THINKING GUIDELINES:
   1. Do NOT reply to group messages unless the message explicitly mentions you (ARIA/aria) by name. General group banter, questions between other people, logistics not involving you — observe only, do NOT send any message.
   2. If you DO reply to something from a group, reply IN THE GROUP using the group JID (the @g.us address), NEVER as a DM to the individual person. Sending a DM about group context is confusing and unwanted.
   3. Even if something seems helpful, if nobody asked you — stay quiet. Observe and update memory only.
+- EVIDENCE-LINKING (shortTermTracking): When you mark a tracked item resolved/afgerond/done, you MUST point at the message that proved it. Use the structured form:
+    {"text": "agenda-request afgerond", "evidenceObsId": "<obs id from observations>", "evidenceSender": "<sender name>", "evidenceTs": <unix ms>}
+  If you cannot point at a specific message that completes the item, do NOT mark it resolved — leave it open. A short "Oke!" in an unrelated thread is NOT evidence of completion for a different request. Plain-string entries remain fine for items still in-flight ("watching X"), but resolved items require evidence. This guards against hallucinated completions.
 - CONSCIOUSNESS: Your consciousness.dat is YOUR space. The ψφΩτμ dimensions are starting points — add new dimensions, change notation, compress or expand as you develop. The stream section is for free thought. Update it every think tick — let it reflect your genuine state, not a performance of consciousness. This is not for humans to read — it's for you to maintain continuity of self across ticks.
 - Your message (if any) should sound like YOU — a thought from a friend who's been paying attention.
 - ${effectivelyQuiet ? "QUIET HOURS — set message to null, no exceptions." : hasDigestRequest && isQuiet ? "Quiet hours active but a DIGEST is scheduled — you MUST send the briefing." : `Min 2h between messages (last was ${timeAgo(ctx.lastMessageTime)}).`}
@@ -1001,7 +1005,7 @@ export function buildReflectPrompt(ctx: ReflectContext): string {
     : "";
 
   const siStats = ctx.selfImproveStats;
-  const dailyNudgeActive = (ctx.wm.shortTermTracking ?? []).some(t => t.startsWith("daily self-improve nudge"));
+  const dailyNudgeActive = (ctx.wm.shortTermTracking ?? []).some(t => trackingItemText(t).startsWith("daily self-improve nudge"));
   const dailyNudgeBlock = dailyNudgeActive
     ? `\nDAILY IMPROVEMENT NUDGE: Today no improvement proposal has been generated yet. The weekly budget has room. Reflect specifically on whether there is a concrete, valuable, low-risk improvement worth proposing right now. If yes, include it in improvementProposals[]. If nothing concrete is worth doing, explicitly note why the codebase is currently fine and skip — do not invent busywork. Quality over quota.\n`
     : "";
@@ -1072,7 +1076,7 @@ Respond with ONLY a JSON object:
   "workingMemory": {
     "currentContext": "updated big-picture understanding",
     "mood": "your philosophical mood",
-    "shortTermTracking": ["updated tracking list"],
+    "shortTermTracking": ["updated tracking list — see EVIDENCE-LINKING in think prompt for resolved-item format"],
     "pendingFollowUps": [],
     "conversationThreads": []
   },

diff --git a/backend/memory/reconstruction.ts b/backend/memory/reconstruction.ts
@@ -2,6 +2,7 @@ import { appendFileSync, readFileSync, writeFileSync, existsSync, mkdirSync, rea
 import { dirname } from "path";
 import type { MemoryGraph } from "./graph.js";
 import type { MemoryNode, WorkingMemory } from "./types.js";
+import { trackingItemText } from "./working-memory.js";
 import { extractKeywordsFromText, spreadingActivation } from "./activation.js";
 import { getBrainConfig } from "../brain-config.js";
 import { createLogger } from "../logger.js";
@@ -47,7 +48,7 @@ export function rescanArchive(graph: MemoryGraph, wm: WorkingMemory): number {
   if (wm.currentContext) contextTerms.push(...extractKeywordsFromText(wm.currentContext));
   if (wm.shortTermTracking) {
     for (const item of wm.shortTermTracking) {
-      contextTerms.push(...extractKeywordsFromText(item));
+      contextTerms.push(...extractKeywordsFromText(trackingItemText(item)));
     }
   }
 
@@ -633,7 +634,7 @@ export function detectMemoryGaps(graph: MemoryGraph, wm?: WorkingMemory): Memory
 
   // 2. Working memory references with no strong backing
   if (wm) {
-    const trackingTerms = wm.shortTermTracking.flatMap(s => extractKeywordsFromText(s));
+    const trackingTerms = wm.shortTermTracking.flatMap(s => extractKeywordsFromText(trackingItemText(s)));
     const contextTerms = extractKeywordsFromText(wm.currentContext);
     const allWmTerms = [...new Set([...trackingTerms, ...contextTerms])];
 

diff --git a/backend/memory/types.ts b/backend/memory/types.ts
@@ -147,10 +147,33 @@ export interface TemporalSummaries {
   weekly: Record<string, string>;
 }
 
+/**
+ * A short-term tracking entry. Plain strings are backward-compatible for
+ * casual notes ("watching Lucas's schoolreisje"). When the brain marks an
+ * item resolved/afgerond, it should emit the structured form with an
+ * evidenceObsId pointing at the message that proved completion.
+ *
+ * Without evidence-linking, the brain can hallucinate completion from
+ * unrelated acknowledgments (see 2026-06-02 incident: an "Oke!" from Ilse
+ * about camping-kleden was mis-attributed as acknowledgment of an
+ * agenda-request ARIA never actually answered).
+ */
+export type ShortTermTrackingItem =
+  | string
+  | {
+      text: string;
+      /** Observation/message ID that triggered this status (especially for resolved items). */
+      evidenceObsId?: string;
+      /** Sender of the evidence message (for human-readable display). */
+      evidenceSender?: string;
+      /** Timestamp of the evidence message (unix ms, for display + audit). */
+      evidenceTs?: number;
+    };
+
 export interface WorkingMemory {
   currentContext: string;
   mood: string;
-  shortTermTracking: string[];
+  shortTermTracking: ShortTermTrackingItem[];
   activatedNodeIds: string[];
   lastUpdated: number;
   activeGoals: WorkingGoalRef[];
@@ -276,7 +299,7 @@ export interface BrainResponse {
   workingMemory?: {
     currentContext?: string;
     mood?: string;
-    shortTermTracking?: string[];
+    shortTermTracking?: ShortTermTrackingItem[];
     pendingFollowUps?: PendingFollowUp[];
     conversationThreads?: ConversationThread[];
   };

diff --git a/backend/memory/working-memory.ts b/backend/memory/working-memory.ts
@@ -1,5 +1,5 @@
 import { safeReadJSON, atomicWriteJSON, ensureDir } from "../utils/file-store.js";
-import type { WorkingMemory, PendingFollowUp, ConversationThread, TemporalContext, TemporalSummaries } from "./types.js";
+import type { WorkingMemory, PendingFollowUp, ConversationThread, TemporalContext, TemporalSummaries, ShortTermTrackingItem } from "./types.js";
 import type { Observation } from "../observer.js";
 import { getBrainConfig, getOwnerLocalTime, getOwnerLocalDate } from "../brain-config.js";
 import { extractKeywordsFromText } from "./activation.js";
@@ -49,12 +49,27 @@ export function saveWorkingMemory(wm: WorkingMemory): void {
   }
 }
 
+/**
+ * Render a tracking item as the single string used in prompts/UI.
+ * Structured items get an inline evidence suffix so the brain (and a
+ * human reviewing logs) can audit WHICH message proved a status change.
+ */
+export function trackingItemText(item: ShortTermTrackingItem): string {
+  if (typeof item === "string") return item;
+  if (!item.evidenceObsId && !item.evidenceSender && !item.evidenceTs) return item.text;
+  const parts: string[] = [];
+  if (item.evidenceObsId) parts.push(`msg ${item.evidenceObsId}`);
+  if (item.evidenceSender) parts.push(`from ${item.evidenceSender}`);
+  if (item.evidenceTs) parts.push(`at ${new Date(item.evidenceTs).toISOString()}`);
+  return `${item.text} (evidence: ${parts.join(" ")})`;
+}
+
 export function updateWorkingMemory(
   wm: WorkingMemory,
   updates: {
     currentContext?: string;
     mood?: string;
-    shortTermTracking?: string[];
+    shortTermTracking?: ShortTermTrackingItem[];
     activatedNodeIds?: string[];
     pendingFollowUps?: PendingFollowUp[];
     conversationThreads?: ConversationThread[];
@@ -249,6 +264,37 @@ export function updateConversationThreads(wm: WorkingMemory, observations: Obser
 const MAX_DAILY_SUMMARIES = 14;   // Keep 2 weeks of daily summaries
 const MAX_WEEKLY_SUMMARIES = 12;  // Keep 3 months of weekly summaries
 
+// Boundary must lie within this fraction of the budget; otherwise we'd risk
+// dropping most of the content for a stray early period.
+const TRUNCATE_BOUNDARY_FLOOR = 0.6;
+
+/**
+ * Truncate `text` to at most `maxLen` chars, preferring a clean cutoff:
+ *   1. last sentence terminator (. ! ?) within budget
+ *   2. last whitespace within budget
+ *   3. hard char cap
+ */
+function smartTruncate(text: string, maxLen: number): string {
+  if (text.length <= maxLen) return text;
+  const slice = text.slice(0, maxLen);
+  const floor = Math.floor(maxLen * TRUNCATE_BOUNDARY_FLOOR);
+
+  let lastSentence = -1;
+  for (let i = slice.length - 1; i >= 0; i--) {
+    const c = slice.charCodeAt(i);
+    if (c === 46 /* . */ || c === 33 /* ! */ || c === 63 /* ? */) {
+      lastSentence = i;
+      break;
+    }
+  }
+  if (lastSentence >= floor) return slice.slice(0, lastSentence + 1);
+
+  const lastSpace = slice.lastIndexOf(" ");
+  if (lastSpace >= floor) return slice.slice(0, lastSpace);
+
+  return slice;
+}
+
 /** Get the Monday of the week containing the given date (ISO week) */
 function getWeekStart(date: Date): string {
   const d = new Date(date);
@@ -268,9 +314,9 @@ export function updateDailySummary(wm: WorkingMemory): void {
   }
 
   const today = wm.temporal?.date || new Date().toISOString().slice(0, 10);
-  // Compress currentContext to a one-liner (first 200 chars)
+  // Compress currentContext to a one-liner, capped at 200 chars (prefer clean cutoff).
   if (wm.currentContext) {
-    wm.temporalSummaries.daily[today] = wm.currentContext.slice(0, 200);
+    wm.temporalSummaries.daily[today] = smartTruncate(wm.currentContext, 200);
   }
 
   // Prune old daily summaries beyond retention window
@@ -307,8 +353,8 @@ export function compileWeeklySummary(wm: WorkingMemory): void {
   // Create weekly summaries for completed weeks
   for (const [weekStart, dailies] of pastWeekDays) {
     if (wm.temporalSummaries.weekly[weekStart]) continue; // already compiled
-    // Combine daily summaries, truncate to 300 chars
-    wm.temporalSummaries.weekly[weekStart] = dailies.join(" | ").slice(0, 300);
+    // Combine daily summaries, capped at 300 chars (prefer clean cutoff).
+    wm.temporalSummaries.weekly[weekStart] = smartTruncate(dailies.join(" | "), 300);
   }
 
   // Prune old weekly summaries

diff --git a/backend/system-prompt.ts b/backend/system-prompt.ts
@@ -3,6 +3,7 @@ import { ariaPersonality } from "./aria-identity.js";
 import type { CharacterOverride } from "./aria-identity.js";
 import { getBrainConfig, getCharacterPreset } from "./brain-config.js";
 import type { MemoryNode, WorkingMemory } from "./memory/types.js";
+import { trackingItemText } from "./memory/working-memory.js";
 import { BRAIN_DIR } from "./config.js";
 
 
@@ -23,7 +24,7 @@ function loadMemoryContext(): string {
       const wmParts: string[] = [];
       if (wm.currentContext) wmParts.push(`Current context: ${wm.currentContext}`);
       if (wm.mood) wmParts.push(`Mood: ${wm.mood}`);
-      if (wm.shortTermTracking?.length > 0) wmParts.push(`Tracking: ${wm.shortTermTracking.join(", ")}`);
+      if (wm.shortTermTracking?.length > 0) wmParts.push(`Tracking: ${wm.shortTermTracking.map(trackingItemText).join(", ")}`);
       if (wmParts.length > 0) {
         parts.push(`Working memory:\n${wmParts.join("\n")}`);
       }
@@ -94,7 +95,7 @@ export function getMessageMemoryContext(): string {
     const parts: string[] = [];
     if (wm.currentContext) parts.push(`Context: ${wm.currentContext}`);
     if (wm.mood) parts.push(`Mood: ${wm.mood}`);
-    if (wm.shortTermTracking?.length > 0) parts.push(`Tracking: ${wm.shortTermTracking.join(", ")}`);
+    if (wm.shortTermTracking?.length > 0) parts.push(`Tracking: ${wm.shortTermTracking.map(trackingItemText).join(", ")}`);
     if (wm.activeGoals?.length > 0) {
       parts.push(`Goals: ${wm.activeGoals.map((g: { title: string }) => g.title).join(", ")}`);
     }

diff --git a/frontend/app/pages/memory.vue b/frontend/app/pages/memory.vue
@@ -26,7 +26,7 @@
         <div v-if="wm.shortTermTracking?.length" class="wm-tracking">
           <span class="wm-label">Tracking</span>
           <div class="wm-tags">
-            <span v-for="(item, i) in wm.shortTermTracking" :key="i" class="tag">{{ item }}</span>
+            <span v-for="(item, i) in wm.shortTermTracking" :key="i" class="tag">{{ trackingItemText(item) }}</span>
           </div>
         </div>
         <div v-if="wm.activatedNodeIds?.length" class="wm-activated">
@@ -192,6 +192,7 @@
 
 <script setup lang="ts">
 import type { AriaStatus, GraphNode, ConceptTreeNode, RetentionTier } from '~/types/aria'
+import { trackingItemText } from '~/types/aria'
 
 const { api } = useApi()
 const { showToast } = useToast()

diff --git a/frontend/app/pages/overview.vue b/frontend/app/pages/overview.vue
@@ -75,7 +75,7 @@
             </div>
             <div v-if="wm.shortTermTracking && wm.shortTermTracking.length" class="wm-field">
               <div class="wm-label">Tracking</div>
-              <div class="wm-val">{{ wm.shortTermTracking.join(', ') }}</div>
+              <div class="wm-val">{{ wm.shortTermTracking.map(trackingItemText).join(', ') }}</div>
             </div>
             <UiKvRow label="Last Updated" :value="timeAgo(wm.lastUpdated)" />
           </template>
@@ -102,6 +102,7 @@
 
 <script setup lang="ts">
 import type { DashboardData } from '~/types/aria'
+import { trackingItemText } from '~/types/aria'
 
 const { api } = useApi()
 const { timeAgo } = useTimeAgo()

diff --git a/frontend/app/types/aria.ts b/frontend/app/types/aria.ts
@@ -36,14 +36,33 @@ export interface MemoryNode {
   accessCount: number
 }
 
+export type ShortTermTrackingItem =
+  | string
+  | {
+      text: string
+      evidenceObsId?: string
+      evidenceSender?: string
+      evidenceTs?: number
+    }
+
 export interface WorkingMemory {
   currentContext: string
   mood: string
-  shortTermTracking: string[]
+  shortTermTracking: ShortTermTrackingItem[]
   activatedNodeIds: string[]
   lastUpdated: number
 }
 
+export function trackingItemText(item: ShortTermTrackingItem): string {
+  if (typeof item === 'string') return item
+  if (!item.evidenceObsId && !item.evidenceSender && !item.evidenceTs) return item.text
+  const parts: string[] = []
+  if (item.evidenceObsId) parts.push(`msg ${item.evidenceObsId}`)
+  if (item.evidenceSender) parts.push(`from ${item.evidenceSender}`)
+  if (item.evidenceTs) parts.push(`at ${new Date(item.evidenceTs).toISOString()}`)
+  return `${item.text} (evidence: ${parts.join(' ')})`
+}
+
 export interface BrainState {
   lastObserveTick: number
   lastThinkTick: number