agent-working-memory 0.17.0 → 0.19.0

package/dist/mcp.js

@@ -50,6 +50,14 @@ try {
     }
 }
 catch { /* No .env file */ }
+// Shared model daemon ON by default for the MCP server (0.19). Measured in 0.18:
+// results byte-identical to in-process, four sessions 4,825 MB → 1,730 MB, and
+// the first recall of every session after the first drops from ~4 s (three
+// models loading into a fresh process) to the warm ~0.6 s. The library default
+// stays off so tests and embedded callers do not spawn daemons; the server
+// entry points opt in here, and AWM_MODEL_DAEMON=0 turns it off.
+if (process.env.AWM_MODEL_DAEMON === undefined)
+    process.env.AWM_MODEL_DAEMON = '1';
 // MCP uses stdout for JSON-RPC. Redirect console.log to stderr so engine
 // startup messages (ConsolidationScheduler, model loading, etc.) don't
 // corrupt the transport. This MUST happen before any engine imports.
@@ -71,6 +79,8 @@ import { DEFAULT_AGENT_CONFIG } from './types/agent.js';
 import { embed, getEmbedder } from './core/embeddings.js';
 import { getReranker } from './core/reranker.js';
 import { getExpander } from './core/query-expander.js';
+import { warmModelDaemon } from './core/ml-worker.js';
+import { isModelDaemonEnabled } from './core/model-daemon.js';
 import { startSidecar } from './hooks/sidecar.js';
 import { initLogger, log, getLogPath } from './core/logger.js';
 import { VERSION } from './version.js';
@@ -81,6 +91,8 @@ import { startLoopLagMonitor } from './core/write-telemetry.js';
 import { buildWhoami, formatWhoami } from './core/whoami.js';
 import { renderTaskEndInvitation, validateRecipeWrite, recipeSlug, getRecipe } from './recipes/index.js';
 import { formatRecallResultLine } from './core/format-recall.js';
+import { newestConflicting } from './core/conflicts.js';
+import { missingValueNote } from './core/asked-value.js';
 import { packRecallByBudget, formatTokenFooter, estimateTokens } from './core/token-budget.js';
 // --- Incognito Mode ---
 // When AWM_INCOGNITO=1, register zero tools. Claude won't see memory tools at all.
@@ -200,15 +212,15 @@ else {
         return 'unclassified';
     }
     // --- Tools ---
-    server.tool('memory_write', `Store a memory. The salience filter decides whether it's worth keeping (active), needs more evidence (staging), or should be discarded.
-
-CALL THIS PROACTIVELY — do not wait to be asked. Write memories when you:
-- Discover something about the codebase, bugs, or architecture
-- Make a decision and want to remember why
-- Encounter and resolve an error
-- Learn a user preference or project pattern
-- Complete a significant piece of work
-
+    server.tool('memory_write', `Store a memory. The salience filter decides whether it's worth keeping (active), needs more evidence (staging), or should be discarded.
+
+CALL THIS PROACTIVELY — do not wait to be asked. Write memories when you:
+- Discover something about the codebase, bugs, or architecture
+- Make a decision and want to remember why
+- Encounter and resolve an error
+- Learn a user preference or project pattern
+- Complete a significant piece of work
+
 The concept should be a short label (3-8 words). The content should be the full detail.`, {
         concept: z.string().describe('Short label for this memory (3-8 words)'),
         content: z.string().describe('Full detail of what was learned'),
@@ -230,6 +242,8 @@ The concept should be a short label (3-8 words). The content should be the full
             .describe('Memory type: episodic (events/incidents), semantic (facts/decisions), procedural (how-to/steps). Auto-classified if omitted.'),
         supersedes: z.string().optional()
             .describe('ID of an older memory this one replaces. The old memory is down-ranked, not deleted.'),
+        correction: z.boolean().optional()
+            .describe('This write corrects a fact AWM already holds and you do not have its id. AWM supersedes the closest active memory about the same thing, stores the correction active at no less than that memory\'s confidence, and says what it replaced. Use it whenever a value changed (a time, a price, a status, a decision); written as a plain observation, a changed fact merges into or ranks below the stale version. If nothing close enough exists the write is stored as new and the reply says so. An explicit supersedes outranks this.'),
         // --- Agent-provided metadata (stored as searchable tags) ---
         project: z.string().optional()
             .describe('Project context (e.g., "EquiHub", "AWM"). Becomes a searchable tag.'),
@@ -307,11 +321,15 @@ The concept should be a short label (3-8 words). The content should be the full
             eventType: params.event_type,
             surprise: params.surprise,
             decisionMade: params.decision_made,
+            // Passed through so the salience filter sees a declared decision, not just an
+            // `intent=` tag. The mapping lives in write-pipeline.ts; see the measurement there.
+            intent: params.intent,
             causalDepth: params.causal_depth,
             resolutionEffort: params.resolution_effort,
             memoryClass: params.memory_class,
             memoryType,
             supersedes: params.supersedes,
+            correction: params.correction,
             originClass: params.origin_class,
             writerSession: params.session_id,
             recipeId: params.recipe_id,
@@ -336,9 +354,9 @@ The concept should be a short label (3-8 words). The content should be the full
         const salience = result.salience; // create/supersede always have salience
         const novelty = result.noveltyResult.novelty;
         const isLowSalience = salience.disposition === 'discard';
-        // Decision propagation: when decision_made=true and coordination is enabled,
+        // Decision propagation: when decision_made=true or intent='decision', and coordination is enabled,
         // broadcast to coord_decisions so other agents can discover it
-        if (params.decision_made && coordDb) {
+        if ((params.decision_made || params.intent === 'decision') && coordDb) {
             try {
                 const agent = coordDb.prepare(`SELECT id, current_task FROM coord_agents WHERE name = ? AND status != 'dead' ORDER BY last_seen DESC LIMIT 1`).get(AGENT_ID);
                 if (agent) {
@@ -349,29 +367,40 @@ The concept should be a short label (3-8 words). The content should be the full
         }
         const logDisposition = isLowSalience ? 'low-salience' : salience.disposition;
         log(AGENT_ID, `write:${logDisposition}`, `"${params.concept}" salience=${salience.score.toFixed(2)} novelty=${novelty.toFixed(1)} id=${engram.id}`);
+        // 2026-10-19: say what a correction replaced, or that it replaced nothing.
+        const correctionNote = params.correction
+            ? (result.correctionOf
+                ? `\nCorrected "${result.correctionOf.concept}" (${result.correctionOf.id}) — it is now superseded by this memory.`
+                : `\nNOTE: no existing memory was close enough to correct, so this was stored as new. If you know which memory it replaces, write again with supersedes: '<id>'.`)
+            : '';
+        const trustNote = result.trustBlocked
+            ? `\nNOTE: stored as its own memory, not merged into ${result.trustBlocked.matchedId} — this write is less trusted than that memory (src/core/trust.ts), so it must not replace or be merged into it.`
+            : '';
         return {
             content: [{
                     type: 'text',
                     text: `Stored (${salience.disposition}) "${params.concept}" [${salience.score.toFixed(2)}]\nID: ${engram.id}`
+                        + correctionNote
+                        + trustNote
                         + (isLowSalience
                             ? `\nNOTE: low salience — this memory is kept but demoted and may fade first. If it MUST survive and be recallable, retry with memory_class: 'canonical'. (Discards are audited: reason codes ${JSON.stringify(salience.reasonCodes.slice(0, 4))})`
                             : ''),
                 }],
         };
     });
-    server.tool('memory_recall', `Recall memories relevant to a query. Uses cognitive activation — not keyword search.
-
-ALWAYS call this when:
-- Starting work on a project or topic (recall what you know)
-- Debugging (recall similar errors and solutions)
-- Making decisions (recall past decisions and outcomes)
-- The user mentions a topic you might have stored memories about
-
-Accepts either "query" or "context" parameter — both work identically.
+    server.tool('memory_recall', `Recall memories relevant to a query. Uses cognitive activation — not keyword search.
+
+ALWAYS call this when:
+- Starting work on a project or topic (recall what you know)
+- Debugging (recall similar errors and solutions)
+- Making decisions (recall past decisions and outcomes)
+- The user mentions a topic you might have stored memories about
+
+Accepts either "query" or "context" parameter — both work identically.
 Returns the most relevant memories ranked by text relevance, temporal recency, and associative strength.`, {
         query: z.string().optional().describe('What to search for — describe the situation, question, or topic'),
         context: z.string().optional().describe('Alias for query (either works)'),
-        limit: z.number().optional().default(3).describe('Max memories to return (default 3). Measured on 400 real-store probes: k=3 is the only token-POSITIVE setting (+115 tok/recall, against -680 at k=5) and is 41% faster at p50, because `limit` sizes the cross-encoder rerank pool (min(limit*8, 40) — so 24 at k=3 against 40 at k=5) and rerank is 83% of warm recall time. It costs ~0.5pp success@1 versus k=5 — about 3 queries in 400, inside noise. Raise it explicitly when you want breadth for orientation rather than a specific fact; that is the case this default deliberately does not optimize for.'),
+        limit: z.number().optional().default(3).describe('Max memories to return (default 3). Measured on 400 real-store probes: k=3 is the only token-POSITIVE setting (+115 tok/recall, against -680 at k=5) and is 41% faster at p50, because `limit` sizes the cross-encoder rerank pool (max(limit*4, 40), so 40 at k=3 and at k=5 before the pool filter trims it, 24 at the median) and rerank is 83% of warm recall time. It costs ~0.5pp success@1 versus k=5 — about 3 queries in 400, inside noise. Raise it explicitly when you want breadth for orientation rather than a specific fact; that is the case this default deliberately does not optimize for.'),
         min_score: z.number().optional().default(0.05).describe('Minimum relevance score (default 0.05)'),
         include_staging: z.boolean().optional().default(false).describe('Include weak/unconfirmed memories?'),
         use_reranker: z.boolean().optional().default(true).describe('Use cross-encoder re-ranking for better relevance (default true)'),
@@ -458,8 +487,23 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
         // WHOLE response and not just the result lines. FOOTER_ALLOWANCE is a
         // fixed upper bound on the accounting line (measured at ~15-25 tokens; the
         // unit test asserts it stays under 40).
+        // 2026-10-19: recall confidence on every reply (it was printed only on
+        // abstention, so a caller could not tell a clear winner from a toss-up),
+        // and one line when the returned set holds two versions of a fact — see
+        // src/core/conflicts.ts. Both are reserved from the budget like the footer.
+        const topConfidence = results[0]?.confidence;
+        const confidenceFooter = topConfidence !== undefined ? `\n[confidence: ${topConfidence.toFixed(3)}]` : '';
+        const marked = results.filter(r => (r.conflictsWith ?? []).length > 0);
+        const newest = newestConflicting(results, results.map(r => r.conflictsWith ?? []));
+        const conflictFooter = marked.length > 0
+            ? `\n⚠ ${marked.length} of these memories disagree with each other${newest ? ` — the newest is ${newest}` : ''}. Confirm against the source before stating a value; supersede the stale one when you know which it is.`
+            : '';
+        // The near-miss (2026-10-19): memories plainly about the subject, none holding
+        // the kind of value asked for. That is the context shape that invites a made-up
+        // number; say so. See src/core/asked-value.ts.
+        const missingFooter = missingValueNote(queryText ?? '', results);
         const FOOTER_ALLOWANCE = 40;
-        const reserved = estimateTokens(peerSuffix) + FOOTER_ALLOWANCE;
+        const reserved = estimateTokens(peerSuffix) + estimateTokens(confidenceFooter) + estimateTokens(conflictFooter) + estimateTokens(missingFooter) + FOOTER_ALLOWANCE;
         const packed = packRecallByBudget(results, formatRecallResultLine, params.max_tokens, reserved);
         // A budget too small to admit even the best match must not look like "no
         // memories found" — that is a different fact and would send the caller off
@@ -483,12 +527,12 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
         return {
             content: [{
                     type: 'text',
-                    text: packed.lines.join('\n') + peerSuffix + formatTokenFooter(packed, params.max_tokens) + evFooter,
+                    text: packed.lines.join('\n') + peerSuffix + conflictFooter + missingFooter + formatTokenFooter(packed, params.max_tokens) + confidenceFooter + evFooter,
                 }],
         };
     });
-    server.tool('memory_feedback', `Report whether a recalled memory was actually useful. This updates the memory's confidence score — useful memories become stronger, useless ones weaken.
-
+    server.tool('memory_feedback', `Report whether a recalled memory was actually useful. This updates the memory's confidence score — useful memories become stronger, useless ones weaken.
+
 Always call this after using a recalled memory so the system learns what's valuable.`, {
         engram_id: z.string().describe('ID of the memory (from memory_recall results)'),
         useful: z.boolean().describe('Was this memory actually helpful?'),
@@ -519,8 +563,8 @@ Always call this after using a recalled memory so the system learns what's valua
                 }],
         };
     });
-    server.tool('memory_retract', `Retract a memory that turned out to be wrong. Creates a correction and reduces confidence of related memories.
-
+    server.tool('memory_retract', `Retract a memory that turned out to be wrong. Creates a correction and reduces confidence of related memories.
+
 Use this when you discover a memory contains incorrect information.`, {
         engram_id: z.string().describe('ID of the wrong memory'),
         reason: z.string().describe('Why is this memory wrong?'),
@@ -544,13 +588,13 @@ Use this when you discover a memory contains incorrect information.`, {
                 }],
         };
     });
-    server.tool('memory_supersede', `Replace an outdated memory with a newer one. Unlike retraction (which marks memories as wrong), supersession marks the old memory as outdated but historically correct.
-
-Use this when:
-- A status or count has changed (e.g., "5 reviews done" → "7 reviews done")
-- Architecture or infrastructure evolved (e.g., "two-repo model" → "three-repo model")
-- A schedule or plan was updated
-
+    server.tool('memory_supersede', `Replace an outdated memory with a newer one. Unlike retraction (which marks memories as wrong), supersession marks the old memory as outdated but historically correct.
+
+Use this when:
+- A status or count has changed (e.g., "5 reviews done" → "7 reviews done")
+- Architecture or infrastructure evolved (e.g., "two-repo model" → "three-repo model")
+- A schedule or plan was updated
+
 The old memory stays in the database (searchable for history) but is heavily down-ranked in recall so the current version dominates.`, {
         old_engram_id: z.string().describe('ID of the outdated memory (from memory_recall results, or memory_write\'s own response if you just wrote it)'),
         new_engram_id: z.string().describe('ID of the replacement memory'),
@@ -581,7 +625,7 @@ The old memory stays in the database (searchable for history) but is heavily dow
         const info = await buildWhoami(store, AGENT_ID, 'mcp', sidecarHandle ? sidecarHandle.boundPort() : undefined);
         return { content: [{ type: 'text', text: formatWhoami(info) }] };
     });
-    server.tool('memory_stats', `Get memory health stats — how many memories, confidence levels, association count, and system performance.
+    server.tool('memory_stats', `Get memory health stats — how many memories, confidence levels, association count, and system performance.
 Also shows the activity log path so the user can tail it to see what's happening.`, {}, async () => {
         const metrics = await evalEngine.computeMetrics(AGENT_ID);
         const checkpoint = await store.getCheckpoint(AGENT_ID);
@@ -623,13 +667,13 @@ Also shows the activity log path so the user can tail it to see what's happening
         };
     });
     // --- Checkpointing Tools ---
-    server.tool('memory_checkpoint', `Save your current execution state so you can recover after context compaction.
-
-ALWAYS call this before:
-- Long operations (multi-file generation, large refactors, overnight work)
-- Anything that might fill the context window
-- Switching to a different task
-
+    server.tool('memory_checkpoint', `Save your current execution state so you can recover after context compaction.
+
+ALWAYS call this before:
+- Long operations (multi-file generation, large refactors, overnight work)
+- Anything that might fill the context window
+- Switching to a different task
+
 Also call periodically during long sessions to avoid losing state. The state is saved per-agent and overwrites any previous checkpoint.`, {
         current_task: z.string().describe('What you are currently working on'),
         decisions: z.array(z.string()).optional().default([])
@@ -663,14 +707,14 @@ Also call periodically during long sessions to avoid losing state. The state is
                 }],
         };
     });
-    server.tool('memory_restore', `Restore your previous execution state after context compaction or at session start.
-
-Returns:
-- Your saved execution state (task, decisions, next steps, files)
-- Recently recalled memories for context
-- Your last write for continuity
-- How long you were idle
-
+    server.tool('memory_restore', `Restore your previous execution state after context compaction or at session start.
+
+Returns:
+- Your saved execution state (task, decisions, next steps, files)
+- Recently recalled memories for context
+- Your last write for continuity
+- How long you were idle
+
 Use this at the start of every session or after compaction to pick up where you left off.`, {}, async () => {
         const checkpoint = await store.getCheckpoint(AGENT_ID);
         // Cold-store nudge: an empty store means the agent has nothing to recall — offer to warm-start.
@@ -793,9 +837,9 @@ Use this at the start of every session or after compaction to pick up where you
         if (coordDb) {
             try {
                 const myAgent = coordDb.prepare(`SELECT id FROM coord_agents WHERE name = ? AND status != 'dead' ORDER BY last_seen DESC LIMIT 1`).get(AGENT_ID);
-                const peerDecisions = coordDb.prepare(`SELECT d.summary, a.name AS author_name, d.created_at
-           FROM coord_decisions d JOIN coord_agents a ON d.author_id = a.id
-           WHERE d.author_id != ? AND d.created_at > datetime('now', '-30 minutes')
+                const peerDecisions = coordDb.prepare(`SELECT d.summary, a.name AS author_name, d.created_at
+           FROM coord_decisions d JOIN coord_agents a ON d.author_id = a.id
+           WHERE d.author_id != ? AND d.created_at > datetime('now', '-30 minutes')
            ORDER BY d.created_at DESC LIMIT 10`).all(myAgent?.id ?? '');
                 if (peerDecisions.length > 0) {
                     parts.push(`\n**Peer decisions (last 30 min):**`);
@@ -814,11 +858,11 @@ Use this at the start of every session or after compaction to pick up where you
         };
     });
     // --- Onboarding Tools (warm-start a cold store) ---
-    server.tool('onboard_scan', `Scan a project's documentation + repository and return CANDIDATE memories to seed a cold store.
-
-Use this when the store is empty / you're new to a project. The scan is deterministic
-(real file contents, not guesses) — YOUR job is to refine the candidates into atomic,
-recall-shaped memories, run the interview (onboard_questions), confirm with the user, then
+    server.tool('onboard_scan', `Scan a project's documentation + repository and return CANDIDATE memories to seed a cold store.
+
+Use this when the store is empty / you're new to a project. The scan is deterministic
+(real file contents, not guesses) — YOUR job is to refine the candidates into atomic,
+recall-shaped memories, run the interview (onboard_questions), confirm with the user, then
 save the good ones with memory_write (memory_class="canonical"). Nothing is saved by this tool.`, {
         docs: z.array(z.string()).optional()
             .describe('Doc files/dirs to scan (Markdown/text). Defaults to the repo (or cwd).'),
@@ -843,18 +887,18 @@ save the good ones with memory_write (memory_class="canonical"). Nothing is save
         ].join('\n');
         return { content: [{ type: 'text', text }] };
     });
-    server.tool('onboard_questions', `Return the onboarding interview questions. Ask the user ONE at a time, starting with the
+    server.tool('onboard_questions', `Return the onboarding interview questions. Ask the user ONE at a time, starting with the
 goal of the memory system, and ask follow-ups for clarity. Turn each answer into a canonical memory.`, {}, async () => ({
         content: [{ type: 'text', text: INTERVIEW_QUESTIONS.map((q, i) => `${i + 1}. ${q}`).join('\n') }],
     }));
     // --- Task Management Tools ---
-    server.tool('memory_task_add', `Create a task that you need to come back to. Tasks are memories with status and priority tracking.
-
-Use this when:
-- You identify work that needs doing but can't do it right now
-- The user mentions something to do later
-- You want to park a sub-task while focusing on something more urgent
-
+    server.tool('memory_task_add', `Create a task that you need to come back to. Tasks are memories with status and priority tracking.
+
+Use this when:
+- You identify work that needs doing but can't do it right now
+- The user mentions something to do later
+- You want to park a sub-task while focusing on something more urgent
+
 Tasks automatically get high salience so they won't be discarded.`, {
         concept: z.string().describe('Short task title (3-10 words)'),
         content: z.string().describe('Full task description — what needs doing, context, acceptance criteria'),
@@ -894,11 +938,11 @@ Tasks automatically get high salience so they won't be discarded.`, {
                 }],
         };
     });
-    server.tool('memory_task_update', `Update a task's status or priority. Use this to:
-- Start working on a task (open → in_progress)
-- Mark a task done (→ done)
-- Block a task on another (→ blocked)
-- Reprioritize (change priority)
+    server.tool('memory_task_update', `Update a task's status or priority. Use this to:
+- Start working on a task (open → in_progress)
+- Mark a task done (→ done)
+- Block a task on another (→ blocked)
+- Reprioritize (change priority)
 - Unblock a task (clear blocked_by)`, {
         task_id: z.string().describe('ID of the task to update'),
         status: z.enum(['open', 'in_progress', 'blocked', 'done']).optional()
@@ -928,8 +972,8 @@ Tasks automatically get high salience so they won't be discarded.`, {
                 }],
         };
     });
-    server.tool('memory_task_list', `List tasks with optional status filter. Shows tasks ordered by priority (urgent first).
-
+    server.tool('memory_task_list', `List tasks with optional status filter. Shows tasks ordered by priority (urgent first).
+
 Use at the start of a session to see what's pending, or to check blocked/done tasks.`, {
         status: z.enum(['open', 'in_progress', 'blocked', 'done']).optional()
             .describe('Filter by status (omit to see all active tasks)'),
@@ -955,10 +999,10 @@ Use at the start of a session to see what's pending, or to check blocked/done ta
                 }],
         };
     });
-    server.tool('memory_task_next', `Get the single most important task to work on next.
-
-Prioritizes: in_progress tasks first (finish what you started), then by priority level, then oldest first. Skips blocked and done tasks.
-
+    server.tool('memory_task_next', `Get the single most important task to work on next.
+
+Prioritizes: in_progress tasks first (finish what you started), then by priority level, then oldest first. Skips blocked and done tasks.
+
 Use this when you finish a task or need to decide what to do next.`, {}, async () => {
         const next = await store.getNextTask(AGENT_ID);
         if (!next) {
@@ -974,13 +1018,13 @@ Use this when you finish a task or need to decide what to do next.`, {}, async (
         };
     });
     // --- Task Bracket Tools ---
-    server.tool('memory_task_begin', `Signal that you're starting a significant task. Auto-checkpoints current state and recalls relevant memories.
-
-CALL THIS when starting:
-- A multi-step operation (doc generation, large refactor, migration)
-- Work on a new topic or project area
-- Anything that might fill the context window
-
+    server.tool('memory_task_begin', `Signal that you're starting a significant task. Auto-checkpoints current state and recalls relevant memories.
+
+CALL THIS when starting:
+- A multi-step operation (doc generation, large refactor, migration)
+- Work on a new topic or project area
+- Anything that might fill the context window
+
 This ensures your state is saved before you start, and primes recall with relevant context.`, {
         topic: z.string().describe('What task are you starting? (3-15 words)'),
         files: z.array(z.string()).optional().default([])
@@ -1032,13 +1076,13 @@ This ensures your state is saved before you start, and primes recall with releva
                 }],
         };
     });
-    server.tool('memory_task_end', `Signal that you've finished a significant task. Writes a summary memory and auto-checkpoints.
-
-CALL THIS when you finish:
-- A multi-step operation
-- Before switching to a different topic
-- At the end of a work session
-
+    server.tool('memory_task_end', `Signal that you've finished a significant task. Writes a summary memory and auto-checkpoints.
+
+CALL THIS when you finish:
+- A multi-step operation
+- Before switching to a different topic
+- At the end of a work session
+
 This captures what was accomplished so future sessions can recall it.`, {
         summary: z.string().describe('What was accomplished? Include key outcomes, decisions, and any issues.'),
         tags: z.array(z.string()).optional().default([])
@@ -1114,14 +1158,14 @@ This captures what was accomplished so future sessions can recall it.`, {
                 }],
         };
     });
-    server.tool('compress_output', `Compress a STRUCTURED tool output (JSON object/array, query rows, log records) into TOON —
-a compact, schema-aware tabular encoding — before putting it in your context. Cuts ~50-65%
-of the tokens on uniform arrays at zero comprehension cost (validated: models read TOON as
-accurately as JSON). Use this on large tool results you need to keep in context.
-
-Output-only and safe: it never changes the data. Non-JSON / prose is returned unchanged.
-TOON is only emitted when it reproduces the input exactly (self-verified round-trip);
-otherwise you get plain JSON back. When compressed, you also get a 'ref' — call
+    server.tool('compress_output', `Compress a STRUCTURED tool output (JSON object/array, query rows, log records) into TOON —
+a compact, schema-aware tabular encoding — before putting it in your context. Cuts ~50-65%
+of the tokens on uniform arrays at zero comprehension cost (validated: models read TOON as
+accurately as JSON). Use this on large tool results you need to keep in context.
+
+Output-only and safe: it never changes the data. Non-JSON / prose is returned unchanged.
+TOON is only emitted when it reproduces the input exactly (self-verified round-trip);
+otherwise you get plain JSON back. When compressed, you also get a 'ref' — call
 retrieve_original(ref) to get the verbatim source back if you ever need it.`, {
         output: z.string().describe('The tool output to compress — JSON text (preferred) or any string. Non-JSON is returned unchanged.'),
         min_saving_chars: z.number().optional().describe('Only emit TOON if it saves at least this many characters (default 40).'),
@@ -1135,8 +1179,8 @@ retrieve_original(ref) to get the verbatim source back if you ever need it.`, {
             content: [{ type: 'text', text: header + r.text }],
         };
     });
-    server.tool('retrieve_original', `Retrieve the verbatim original text for a 'ref' returned by compress_output. Use this when
-you need the exact, uncompressed source (e.g. to pass it to another tool unchanged). Returns
+    server.tool('retrieve_original', `Retrieve the verbatim original text for a 'ref' returned by compress_output. Use this when
+you need the exact, uncompressed source (e.g. to pass it to another tool unchanged). Returns
 an error if the ref has expired (originals are kept for the most recent compressions only).`, {
         ref: z.string().describe('The ref handle returned by compress_output (e.g. "awm_orig_12").'),
     }, async (params) => {
@@ -1184,11 +1228,14 @@ an error if the ref has expired (originals are kept for the most recent compress
                             ? r.engram.createdAt.toISOString()
                             : r.engram.createdAt,
                         memoryClass: r.engram.memoryClass,
+                        validFrom: r.engram.validFrom,
                         validTo: r.engram.validTo,
+                        supersededBy: r.engram.supersededBy,
                     },
                     score: r.score,
                     summary: r.summary,
                     confidence: r.confidence,
+                    conflictsWith: r.conflictsWith,
                 }));
             },
             onConsolidate: async (agentId, reason) => {
@@ -1206,9 +1253,15 @@ an error if the ref has expired (originals are kept for the most recent compress
         // All output is stderr-safe (stdout carries JSON-RPC frames).
         // Escape hatch: AWM_NO_EAGER_WARM=1 restores lazy loading.
         if (process.env.AWM_NO_EAGER_WARM !== '1') {
-            getEmbedder().catch(err => console.error('Embedding model unavailable:', err.message));
-            getReranker().catch(err => console.error('Reranker model unavailable:', err.message));
-            getExpander().catch(err => console.error('Query expander model unavailable:', err.message));
+            // With the shared model daemon on, the daemon loads them and this process loads none.
+            if (isModelDaemonEnabled()) {
+                void warmModelDaemon();
+            }
+            else {
+                getEmbedder().catch(err => console.error('Embedding model unavailable:', err.message));
+                getReranker().catch(err => console.error('Reranker model unavailable:', err.message));
+                getExpander().catch(err => console.error('Query expander model unavailable:', err.message));
+            }
             if (BACKEND === 'sqlite') {
                 setImmediate(() => {
                     try {