ZBBS-WORK-391: dream_source=notes — dream from curated notes (MEM-137)

migrations/MEM-137-dream-source_down.sql

@@ -0,0 +1,5 @@
+-- MEM-137 down — remove the per-agent dream_source column.
+-- Dropping the column also drops its CHECK constraint.
+
+ALTER TABLE agent_configuration
+    DROP COLUMN dream_source;

migrations/MEM-137-dream-source_up.sql

@@ -0,0 +1,26 @@
+-- MEM-137 — per-agent dream source: conversation logs (default) or curated
+-- notes (ZBBS-WORK-391).
+--
+-- Motivating case: a companion-mode account whose memory model is hand
+-- curation — it has zero conversations/* notes, so dream_mode=companion is a
+-- silent no-op (the dream cron sources exclusively from conversations/%).
+-- Its substance lives in curated notes (identity docs, journals, session
+-- summaries) keyed by updated_at.
+--
+-- dream_source selects where the cron reads its raw material:
+--   'conversation' — conversations/% windowed by created_at (existing
+--                    behavior, the default; nothing changes for existing
+--                    agents).
+--   'notes'        — all namespace notes EXCEPT the pipeline's own outputs
+--                    (conversations/dreams/context/learnings prefixes),
+--                    windowed by updated_at. An edit to a curated note
+--                    re-enters it as fresh dream material on the next run.
+--
+-- TEXT + CHECK rather than an enum type: two values, and a CHECK is cheaper
+-- to extend or drop than ALTER TYPE ... ADD VALUE (which can't run inside a
+-- transaction on older PG). dream_mode predates this convention.
+
+ALTER TABLE agent_configuration
+    ADD COLUMN dream_source TEXT NOT NULL DEFAULT 'conversation'
+    CONSTRAINT agent_configuration_dream_source_check
+    CHECK (dream_source IN ('conversation', 'notes'));

node/api/src/routes/admin.js

@@ -377,7 +377,7 @@ router.post('/admin/agents', requirePerm('agents', 'read'), adminRoute('agents-l
     const visibleIds = await getVisibleActorIds(req.actorId);
     // Subqueries for visibility and VA access summaries shown in the agent list
     let sql = `SELECT s.agent, s.actor_id, s.status, s.last_seen, s.passphrase_rotated_at, s.registered_at, s.provider, s.model, s.virtual, s.personality, s.active_since,
-                s.cost_budget_daily, s.cost_budget_monthly, s.cache_prompts, s.learning_enabled, s.max_tokens, s.temperature, s.dream_mode, s.storage_quota, ac.configuration,
+                s.cost_budget_daily, s.cost_budget_monthly, s.cache_prompts, s.learning_enabled, s.max_tokens, s.temperature, s.dream_mode, ac.dream_source, s.storage_quota, ac.configuration,
                 a.realms,
                 COALESCE(vis.summary, 'self only') AS visibility_summary,
                 va.summary AS va_access_summary
@@ -1519,7 +1519,7 @@ router.post('/admin/actors/create', requirePerm('agents', 'write'), adminRoute('
     const { name, provider, model, welcome_template_id, welcome_note_template_id, virtual: isVirtual, personality,
             cost_budget_daily, cost_budget_monthly,
             cache_prompts, learning_enabled, max_tokens, temperature, configuration,
-            ui_access, password, dream_mode } = req.body;
+            ui_access, password, dream_mode, dream_source } = req.body;
 
     if (!name || !name.trim()) {
         return res.status(400).json({
@@ -1607,8 +1607,8 @@ router.post('/admin/actors/create', requirePerm('agents', 'write'), adminRoute('
 
         // Create agent configuration
         await client.query(
-            `INSERT INTO agent_configuration (actor_id, provider, model, virtual, personality, cost_budget_daily, cost_budget_monthly, cache_prompts, learning_enabled, max_tokens, temperature, configuration, dream_mode)
-             VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13)`,
+            `INSERT INTO agent_configuration (actor_id, provider, model, virtual, personality, cost_budget_daily, cost_budget_monthly, cache_prompts, learning_enabled, max_tokens, temperature, configuration, dream_mode, dream_source)
+             VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14)`,
             [actorId, provider || null, model || null,
              isVirtual === true, personality || null,
              parseCostBudget(cost_budget_daily, 'cost_budget_daily'),
@@ -1617,7 +1617,8 @@ router.post('/admin/actors/create', requirePerm('agents', 'write'), adminRoute('
              max_tokens != null ? safeInt(max_tokens) : null,
              temperature != null ? parseFloat(temperature) : null,
              configuration ? JSON.stringify(sanitizeAgentConfiguration(configuration)) : null,
-             ['none', 'companion', 'technical', 'sim'].includes(dream_mode) ? dream_mode : 'none']
+             ['none', 'companion', 'technical', 'sim'].includes(dream_mode) ? dream_mode : 'none',
+             ['conversation', 'notes'].includes(dream_source) ? dream_source : 'conversation']
         );
 
         // Grant all MCP permissions (full tool access) — non-virtual agents only
@@ -1739,7 +1740,7 @@ router.post('/admin/agents/read', requirePerm('agents', 'read'), adminRoute('age
                 agc.cache_prompts, agc.learning_enabled, agc.max_tokens, agc.temperature,
                 agc.cost_budget_daily, agc.cost_budget_monthly, agc.storage_quota,
                 agc.api_key IS NOT NULL AS has_api_key,
-                agc.dream_mode, ac.realms
+                agc.dream_mode, agc.dream_source, ac.realms
          FROM agent_configuration agc
          JOIN actors ac ON ac.id = agc.actor_id
          WHERE agc.actor_id = $1`,
@@ -1796,7 +1797,7 @@ router.post('/admin/agents/update', requirePerm('agents', 'write'), adminRoute('
     const agent = sanitize.agentName(req.body.agent);
     const { personality, api_key, configuration, provider, model,
             cost_budget_daily, cost_budget_monthly, storage_quota,
-            cache_prompts, learning_enabled, max_tokens, temperature, dream_mode, realms } = req.body;
+            cache_prompts, learning_enabled, max_tokens, temperature, dream_mode, dream_source, realms } = req.body;
 
     if (!agent) {
         return res.status(400).json({
@@ -1902,6 +1903,15 @@ router.post('/admin/agents/update', requirePerm('agents', 'write'), adminRoute('
         params.push(dream_mode);
         updates.push(`dream_mode = $${idx++}`);
     }
+    if (dream_source !== undefined) {
+        if (!['conversation', 'notes'].includes(dream_source)) {
+            return res.status(400).json({
+                error: { code: 'BAD_REQUEST', message: 'dream_source must be conversation or notes' }
+            });
+        }
+        params.push(dream_source);
+        updates.push(`dream_source = $${idx++}`);
+    }
 
     // Realms lives on actors table, not agent_configuration — handle separately
     const hasRealmsUpdate = Array.isArray(realms);

node/api/src/services/dream.js

@@ -1,5 +1,6 @@
 // Dream processing — nightly conversation log analysis.
-// Reads conversation logs uploaded by agents, sends them through a dream
+// Reads conversation logs uploaded by agents (or, for dream_source=notes
+// agents, their curated notes — see MEM-137), sends them through a dream
 // virtual agent (companion or technical), and saves consolidated insights
 // as notes in the agent's namespace.
 
@@ -180,6 +181,20 @@ function prefilterLog(content) {
     return result.join('\n');
 }
 
+// Assemble the notes-mode dream source text (MEM-137). Each curated note
+// gets a header carrying its slug and last-updated date so the dream agent
+// can attribute material to a document, mirroring how conversation logs
+// carry their own timestamps. Rows arrive from the notes-mode source query
+// ordered by updated_at ASC.
+function buildNotesLog(rows) {
+    return rows.map(r => {
+        const updated = r.updated_at instanceof Date
+            ? r.updated_at.toISOString().slice(0, 10)
+            : String(r.updated_at).slice(0, 10);
+        return '## Note: ' + r.slug + ' (updated ' + updated + ')\n\n' + r.content;
+    }).join('\n\n---\n\n');
+}
+
 // Extract unique speakers from conversation logs and group lines by speaker.
 // Handles five formats:
 //   memory-sync uploads:    "[HH:MM speaker] message text"
@@ -505,21 +520,58 @@ async function processDreamChunk(agent, agentNames, chunk) {
     const { dreamAgentName, soulAgentName, peopleAgentName, learningsAgentName } = agentNames;
     const { from, to } = chunk;
     const chunkDateStr = from.toISOString().slice(0, 10);
-
-    const logs = await pool.query(
-        `SELECT slug, content, created_at FROM documents
-         WHERE namespace = $1 AND slug LIKE 'conversations/%' AND deleted_at IS NULL
-         AND created_at > $2 AND created_at <= $3
-         ORDER BY created_at ASC`,
-        [agent.name, from, to]
-    );
+    const notesMode = agent.dream_source === 'notes';
+
+    let logs;
+    if (notesMode) {
+        // Notes-sourced dreaming (MEM-137): the raw material is the agent's
+        // own curated notes, windowed by updated_at instead of created_at —
+        // so a later human edit to a note re-enters it as fresh material on
+        // the next run. The prefix exclusions are load-bearing: without them
+        // the cron would dream about its own dreams/soul/learnings output and
+        // feed back on itself (the same spiral that bloated the technical
+        // souls). conversations/% is excluded because dream_source selects
+        // ONE source — agents with real conversation logs use the default.
+        logs = await pool.query(
+            `SELECT slug, content, updated_at FROM documents
+             WHERE namespace = $1 AND deleted_at IS NULL
+             AND slug NOT LIKE 'conversations/%'
+             AND slug NOT LIKE 'dreams/%'
+             AND slug NOT LIKE 'context/%'
+             AND slug NOT LIKE 'learnings/%'
+             AND updated_at > $2 AND updated_at <= $3
+             ORDER BY updated_at ASC`,
+            [agent.name, from, to]
+        );
+    } else {
+        logs = await pool.query(
+            `SELECT slug, content, created_at FROM documents
+             WHERE namespace = $1 AND slug LIKE 'conversations/%' AND deleted_at IS NULL
+             AND created_at > $2 AND created_at <= $3
+             ORDER BY created_at ASC`,
+            [agent.name, from, to]
+        );
+    }
     if (logs.rows.length === 0) {
-        logDream('chunk-no-logs', { agent: agent.name, chunkDate: chunkDateStr });
+        logDream('chunk-no-logs', { agent: agent.name, source: agent.dream_source, chunkDate: chunkDateStr });
         return { skipped: true, reason: 'no logs', chunkDate: chunkDateStr };
     }
 
-    const fullLog = logs.rows.map(r => r.content).join('\n\n---\n\n');
-    const filtered = prefilterLog(fullLog);
+    // In notes mode each note gets a slug+date header so the dream agent can
+    // tell documents apart (conversation logs carry their own timestamps;
+    // curated notes don't). No signal prefilter either: SIGNAL_PATTERNS are
+    // conversational markers ("remember", "don't do that") that deliberate
+    // prose rarely contains — filtering would drop most of the material as
+    // "no signals". Curated notes are already distilled; feed them whole.
+    let fullLog;
+    let filtered;
+    if (notesMode) {
+        fullLog = buildNotesLog(logs.rows);
+        filtered = fullLog;
+    } else {
+        fullLog = logs.rows.map(r => r.content).join('\n\n---\n\n');
+        filtered = prefilterLog(fullLog);
+    }
     if (!filtered) {
         logDream('chunk-no-signals', { agent: agent.name, chunkDate: chunkDateStr, logCount: logs.rows.length });
         return { skipped: true, reason: 'no signals', chunkDate: chunkDateStr, logCount: logs.rows.length };
@@ -528,13 +580,17 @@ async function processDreamChunk(agent, agentNames, chunk) {
     logDream('chunk-processing', {
         agent: agent.name,
         mode: agent.dream_mode,
+        source: agent.dream_source,
         chunkDate: chunkDateStr,
         logCount: logs.rows.length,
         originalSize: fullLog.length,
         filteredSize: filtered.length,
     });
 
-    const userMessage = 'Conversation logs for agent "' + agent.name + '" on ' + chunkDateStr + ':\n\n'
+    const userMessage = (notesMode
+        ? 'Curated notes written or updated by agent "' + agent.name + '" on ' + chunkDateStr
+            + ' (this agent\'s memory lives in hand-curated notes — journals, identity documents, session summaries — rather than conversation logs):\n\n'
+        : 'Conversation logs for agent "' + agent.name + '" on ' + chunkDateStr + ':\n\n')
         + filtered
         + '\n\nAlso provide a brief title summarizing the overarching subject of the day.';
 
@@ -664,7 +720,13 @@ async function processDreamChunk(agent, agentNames, chunk) {
     // dream mode (currently companion and sim). Per-chunk for the same
     // reason soul does: per-day relationship updates compose better than
     // one massive end-of-run pass over weeks of conversation.
-    if (peopleAgentName) {
+    //
+    // Skipped in notes mode: extractSpeakers parses conversation formats
+    // (chat timestamps, discussion lines, sim distiller output) and would
+    // misparse curated prose into junk context/people/* files — e.g. any
+    // "word:" line in a journal becomes a phantom speaker. Notes-sourced
+    // dreaming deliberately writes only dreams/* and context/soul.
+    if (peopleAgentName && !notesMode) {
         try {
             const speakers = extractSpeakers(filtered, agent.name);
             for (const [slug, entry] of speakers) {
@@ -703,7 +765,10 @@ async function processDreamChunk(agent, agentNames, chunk) {
     // for sim agents, where every in-world tick is tool-use and the per-turn
     // extractor is silenced by its !isToolUse gate. Distills the day's
     // filtered conversation into a single learnings note keyed by date.
-    if (learningsAgentName) {
+    // Skipped in notes mode for the same containment reason as people above
+    // (and learnings/% is an excluded source prefix — writing it would feed
+    // the next run's input).
+    if (learningsAgentName && !notesMode) {
         const learningsSlug = 'learnings/' + chunkDateStr + '-sim-day';
         try {
             let existingFile = '';
@@ -795,8 +860,8 @@ async function runDream() {
 
     // Find agents with dream mode enabled
     const agents = await pool.query(
-        `SELECT ac.name, ac.id AS actor_id, agc.dream_mode, agc.last_dream_at,
-                agc.startup_instructions
+        `SELECT ac.name, ac.id AS actor_id, agc.dream_mode, agc.dream_source,
+                agc.last_dream_at, agc.startup_instructions
          FROM agent_configuration agc
          JOIN actors ac ON ac.id = agc.actor_id
          WHERE agc.dream_mode IN ('companion', 'technical', 'sim')`
@@ -840,8 +905,40 @@ async function runDream() {
             // Split the work since last_dream_at into per-UTC-day chunks so an
             // agent that's fallen behind doesn't try to fit weeks of logs into
             // one model call (which is what tripped home with deepseek's 163K
-            // window). First-run agents process the previous 24h.
-            const since = agent.last_dream_at || new Date(Date.now() - 24 * 60 * 60 * 1000);
+            // window). First-run agents process the previous 24h — except in
+            // notes mode, where the first run starts at the earliest curated
+            // note so the soul accretes over the full history in authored
+            // order. (Day-chunks with no notes skip cheaply, so a long span
+            // costs only the days that actually have material.)
+            let since = agent.last_dream_at;
+            if (!since && agent.dream_source === 'notes') {
+                const earliest = await pool.query(
+                    `SELECT MIN(updated_at) AS min_updated FROM documents
+                     WHERE namespace = $1 AND deleted_at IS NULL
+                     AND slug NOT LIKE 'conversations/%'
+                     AND slug NOT LIKE 'dreams/%'
+                     AND slug NOT LIKE 'context/%'
+                     AND slug NOT LIKE 'learnings/%'`,
+                    [agent.name]
+                );
+                if (!earliest.rows[0] || !earliest.rows[0].min_updated) {
+                    logDream('no-notes', { agent: agent.name });
+                    results.push({ agent: agent.name, skipped: true, reason: 'dream_source=notes but namespace has no source notes' });
+                    continue;
+                }
+                // The chunk window is exclusive on `from` (updated_at > from),
+                // so back off 1ms to include the earliest note itself.
+                // Coerce defensively — a custom pg type parser could hand
+                // back a string instead of a Date (same guard buildNotesLog
+                // applies to updated_at).
+                const minUpdated = earliest.rows[0].min_updated instanceof Date
+                    ? earliest.rows[0].min_updated
+                    : new Date(earliest.rows[0].min_updated);
+                since = new Date(minUpdated.getTime() - 1);
+            }
+            if (!since) {
+                since = new Date(Date.now() - 24 * 60 * 60 * 1000);
+            }
             const chunks = computeDailyChunks(since, new Date());
 
             if (chunks.length === 0) {
@@ -967,4 +1064,4 @@ function startDreamScheduler() {
     logDream('scheduler', { message: 'Dream scheduler started', schedule });
 }
 
-module.exports = { runDream, prefilterLog, extractSpeakers, startDreamScheduler, runPersonContextUpdate, findDreamAgent };
+module.exports = { runDream, prefilterLog, extractSpeakers, buildNotesLog, startDreamScheduler, runPersonContextUpdate, findDreamAgent };

node/api/src/services/dream.test.js

@@ -0,0 +1,49 @@
+// Tests for buildNotesLog — the notes-mode (dream_source=notes, MEM-137)
+// source-text assembler in the dream cron. Run with: node --test (from
+// node/api). Uses the built-in node:test runner + node:assert, matching
+// sim-conversation-distiller.test.js.
+//
+// buildNotesLog is the only new pure surface of ZBBS-WORK-391; the sourcing
+// branch and first-run window live in SQL + cron flow and are exercised at
+// deploy (the conversation-mode path is untouched by the change).
+
+const { test } = require('node:test');
+const assert = require('node:assert/strict');
+const { buildNotesLog } = require('./dream');
+
+test('single note gets a slug+date header above its content', () => {
+    const rows = [{
+        slug: 'journal/2026-05-california-trip',
+        content: 'We drove the coast road.',
+        updated_at: new Date('2026-05-18T14:22:00Z'),
+    }];
+    assert.equal(
+        buildNotesLog(rows),
+        '## Note: journal/2026-05-california-trip (updated 2026-05-18)\n\nWe drove the coast road.'
+    );
+});
+
+test('multiple notes are separated and keep their query order', () => {
+    const rows = [
+        { slug: 'core-identity', content: 'A', updated_at: new Date('2026-05-16T08:00:00Z') },
+        { slug: 'thea-core', content: 'B', updated_at: new Date('2026-05-16T19:30:00Z') },
+    ];
+    assert.equal(
+        buildNotesLog(rows),
+        '## Note: core-identity (updated 2026-05-16)\n\nA'
+        + '\n\n---\n\n'
+        + '## Note: thea-core (updated 2026-05-16)\n\nB'
+    );
+});
+
+test('string updated_at (non-Date driver output) still renders the date', () => {
+    const rows = [{
+        slug: 'sirius-identity',
+        content: 'C',
+        updated_at: '2026-06-06T03:15:00.000Z',
+    }];
+    assert.equal(
+        buildNotesLog(rows),
+        '## Note: sirius-identity (updated 2026-06-06)\n\nC'
+    );
+});