@vellumai/assistant 0.7.3 → 0.8.0

package/src/memory/v2/qdrant.ts

@@ -48,21 +48,43 @@ export interface ConceptPagePayload {
 export interface ConceptPageQueryResult {
   slug: string;
   /**
-   * Dense cosine similarity, when the slug appeared in the dense top-`limit`.
-   * `undefined` if the slug only appeared in the sparse channel.
+   * Dense cosine similarity against the page body, when the slug appeared in
+   * the body dense top-`limit`. `undefined` if the slug only appeared in the
+   * sparse channel — or in a summary-side channel.
    */
   denseScore?: number;
   /**
-   * Sparse score, when the slug appeared in the sparse top-`limit`.
-   * `undefined` if the slug only appeared in the dense channel. Lives on a
-   * different scale than `denseScore` — callers must normalize before fusing.
+   * Sparse score against the page body, when the slug appeared in the body
+   * sparse top-`limit`. `undefined` if the slug only appeared in the dense
+   * channel. Lives on a different scale than `denseScore` — callers must
+   * normalize before fusing.
    */
   sparseScore?: number;
+  /**
+   * Dense cosine similarity against the page's frontmatter `summary`, when
+   * the page has a summary embedded and the slug appeared in the summary
+   * dense top-`limit`. `undefined` for pages without a summary embedding —
+   * those fall back to body-only scoring.
+   */
+  summaryDenseScore?: number;
+  /**
+   * Sparse score against the page's frontmatter `summary`, paired with
+   * `summaryDenseScore`. `undefined` for pages without a summary embedding.
+   */
+  summarySparseScore?: number;
 }
 
 let _client: QdrantRestClient | null = null;
 let _collectionReady = false;
-let _collectionReadyPromise: Promise<void> | null = null;
+let _collectionReadyPromise: Promise<{ migrated: boolean }> | null = null;
+
+/**
+ * Named vectors the v2 concept-page collection must expose. Existing
+ * collections that lack any of these get destructively recreated by
+ * `ensureConceptPageCollectionOnce` — see the `migrated` return flag.
+ */
+const REQUIRED_DENSE_VECTORS = ["dense", "summary_dense"] as const;
+const REQUIRED_SPARSE_VECTORS = ["sparse", "summary_sparse"] as const;
 
 /** Lazily create a Qdrant REST client bound to the resolved URL. */
 function getClient(): QdrantRestClient {
@@ -76,16 +98,19 @@ function getClient(): QdrantRestClient {
 }
 
 /**
- * Create the v2 concept-page collection if it does not already exist.
- * Idempotent: a no-op when the collection is already present.
- *
- * Vector layout mirrors `VellumQdrantClient.ensureCollection` — named dense
- * (cosine, configurable size + on-disk) and sparse vectors. The vector size
- * and on-disk flag inherit from `config.memory.qdrant` so v2 stays aligned
- * with the user's existing embedding backend without separate knobs.
+ * Create the v2 concept-page collection if it does not already exist, or
+ * destructively recreate it when the existing schema is missing any of the
+ * required named vectors (see `REQUIRED_DENSE_VECTORS` /
+ * `REQUIRED_SPARSE_VECTORS`). The latter case is signalled to callers via
+ * `{ migrated: true }` so they can enqueue a backfill — pre-#29823
+ * collections lack `summary_dense` / `summary_sparse` and every query
+ * referencing those named vectors fails with HTTP 400 until the collection
+ * is rebuilt. Mirrors `VellumQdrantClient.ensureCollection` for v1.
  */
-export async function ensureConceptPageCollection(): Promise<void> {
-  if (_collectionReady) return;
+export async function ensureConceptPageCollection(): Promise<{
+  migrated: boolean;
+}> {
+  if (_collectionReady) return { migrated: false };
   if (_collectionReadyPromise) return _collectionReadyPromise;
 
   _collectionReadyPromise = ensureConceptPageCollectionOnce().finally(() => {
@@ -94,17 +119,46 @@ export async function ensureConceptPageCollection(): Promise<void> {
   return _collectionReadyPromise;
 }
 
-async function ensureConceptPageCollectionOnce(): Promise<void> {
+async function ensureConceptPageCollectionOnce(): Promise<{
+  migrated: boolean;
+}> {
   const client = getClient();
   const config = getConfig();
   const vectorSize = config.memory.qdrant.vectorSize;
   const onDisk = config.memory.qdrant.onDisk;
 
+  let migrated = false;
+
   try {
     const exists = await client.collectionExists(MEMORY_V2_COLLECTION);
     if (exists.exists) {
-      _collectionReady = true;
-      return;
+      // Assume compatible on probe failure rather than risk a destructive
+      // recreate — mirrors v1's posture in `VellumQdrantClient.ensureCollection`.
+      let info: Awaited<ReturnType<typeof client.getCollection>>;
+      try {
+        info = await client.getCollection(MEMORY_V2_COLLECTION);
+      } catch (err) {
+        log.warn(
+          { err, collection: MEMORY_V2_COLLECTION },
+          "Failed to probe v2 collection schema; assuming compatible",
+        );
+        _collectionReady = true;
+        return { migrated: false };
+      }
+
+      const missing = missingNamedVectors(info);
+      if (missing.length === 0) {
+        _collectionReady = true;
+        return { migrated: false };
+      }
+
+      log.warn(
+        { collection: MEMORY_V2_COLLECTION, missingNamedVectors: missing },
+        "Memory v2 concept-page collection schema drift detected — deleting and recreating; embeddings will be regenerated by background reembed",
+      );
+      await client.deleteCollection(MEMORY_V2_COLLECTION);
+      migrated = true;
+      // Fall through to creation below.
     }
   } catch (err) {
     // Treat "not found"-shaped errors as "needs creation" and fall through.
@@ -124,15 +178,28 @@ async function ensureConceptPageCollectionOnce(): Promise<void> {
           distance: "Cosine",
           on_disk: onDisk,
         },
+        // Optional second dense vector covering the page's frontmatter
+        // `summary`. Pages without a summary store nothing under this name —
+        // Qdrant supports per-point named-vector subsets — so the named-vector
+        // index stays cheap until summaries are populated.
+        summary_dense: {
+          size: vectorSize,
+          distance: "Cosine",
+          on_disk: onDisk,
+        },
       },
       sparse_vectors: {
         sparse: {}, // Qdrant auto-infers sparse vector params
+        summary_sparse: {}, // BM25 sparse vector for the summary
       },
       hnsw_config: {
         on_disk: onDisk,
         m: 16,
         ef_construct: 100,
       },
+      optimizers_config: {
+        default_segment_number: 2,
+      },
       on_disk_payload: onDisk,
     });
   } catch (err) {
@@ -143,7 +210,7 @@ async function ensureConceptPageCollectionOnce(): Promise<void> {
       (err as { status: number }).status === 409
     ) {
       _collectionReady = true;
-      return;
+      return { migrated };
     }
     throw err;
   }
@@ -156,32 +223,86 @@ async function ensureConceptPageCollectionOnce(): Promise<void> {
   });
 
   _collectionReady = true;
+  return { migrated };
+}
+
+/**
+ * Return the names of required named vectors absent from the collection's
+ * current schema. An empty array means the collection is fully migrated.
+ *
+ * If the response shape is unparseable (e.g. Qdrant returns an unexpected
+ * structure) we treat it as "everything is missing" so the caller's drift
+ * branch fires — combined with the `getCollection` try/catch in the caller,
+ * a thrown probe falls back to "assume compatible" while a parsed-but-empty
+ * response triggers the safer recreate.
+ */
+function missingNamedVectors(
+  info: Awaited<ReturnType<QdrantRestClient["getCollection"]>>,
+): string[] {
+  const params = info.config?.params;
+  const dense = params?.vectors;
+  const sparse = (params as { sparse_vectors?: unknown } | undefined)
+    ?.sparse_vectors;
+  const denseNames =
+    dense && typeof dense === "object" && !("size" in dense)
+      ? new Set(Object.keys(dense))
+      : new Set<string>();
+  const sparseNames =
+    sparse && typeof sparse === "object"
+      ? new Set(Object.keys(sparse as Record<string, unknown>))
+      : new Set<string>();
+
+  const missing: string[] = [];
+  for (const name of REQUIRED_DENSE_VECTORS) {
+    if (!denseNames.has(name)) missing.push(name);
+  }
+  for (const name of REQUIRED_SPARSE_VECTORS) {
+    if (!sparseNames.has(name)) missing.push(name);
+  }
+  return missing;
 }
 
 /**
  * Upsert a concept page's dense + sparse embedding. The point ID is derived
  * deterministically from the slug so subsequent calls for the same slug
  * replace the prior point in place rather than accumulating duplicates.
+ *
+ * `summary` is optional — supplied when the page's frontmatter carries a
+ * `summary`, omitted otherwise. Pages without a summary store only the body
+ * vectors and fall back to body-only scoring at query time. The grouped
+ * shape enforces at the type level that summary dense and sparse are
+ * always written together.
  */
 export async function upsertConceptPageEmbedding(params: {
   slug: string;
   dense: number[];
   sparse: SparseEmbedding;
+  summary?: { dense: number[]; sparse: SparseEmbedding };
   updatedAt: number;
 }): Promise<void> {
   await ensureConceptPageCollection();
 
-  const { slug, dense, sparse, updatedAt } = params;
+  const { slug, dense, sparse, summary, updatedAt } = params;
   const client = getClient();
   const pointId = pointIdForSlug(slug);
 
+  // Qdrant lets us upsert any subset of named vectors per point. The summary
+  // entries appear only when the caller passed a `summary` block — pairing
+  // them at the type level keeps query-time fusion symmetric with the body
+  // channels.
+  const vector: Record<string, number[] | SparseEmbedding> = { dense, sparse };
+  if (summary) {
+    vector.summary_dense = summary.dense;
+    vector.summary_sparse = summary.sparse;
+  }
+
   const upsertOnce = () =>
     client.upsert(MEMORY_V2_COLLECTION, {
       wait: true,
       points: [
         {
           id: pointId,
-          vector: { dense, sparse },
+          vector,
           payload: { slug, updated_at: updatedAt },
         },
       ],
@@ -290,6 +411,30 @@ export async function pruneSlugsWithPrefixExcept(
   }
 }
 
+/**
+ * Approximate count of points in the v2 concept-page collection. Used by the
+ * daemon-startup rebuild hook to detect "collection exists but empty" — the
+ * crash-mid-rebuild recovery case where a prior boot dropped + recreated the
+ * collection but died before reembed completed. Returns `0` if the collection
+ * does not exist or the count call fails (treated as "needs reembed" by the
+ * caller).
+ */
+export async function countConceptPagePoints(): Promise<number> {
+  await ensureConceptPageCollection();
+  try {
+    const result = await getClient().count(MEMORY_V2_COLLECTION, {
+      exact: false,
+    });
+    return result.count;
+  } catch (err) {
+    log.warn(
+      { err, collection: MEMORY_V2_COLLECTION },
+      "Failed to count v2 concept-page collection — treating as empty",
+    );
+    return 0;
+  }
+}
+
 /**
  * Best-effort delete of the legacy `memory_v2_skills` Qdrant collection. Skill
  * embeddings now live alongside concept pages in `memory_v2_concept_pages`
@@ -319,9 +464,15 @@ export async function dropLegacySkillsCollection(): Promise<void> {
  * a normalized weighted-sum — because RRF would discard the score magnitudes
  * the activation formula needs.
  *
+ * Four channels are queried concurrently: body dense, body sparse, summary
+ * dense, summary sparse. The summary channels only return hits for pages whose
+ * frontmatter carries a `summary` (and therefore stored `summary_dense` /
+ * `summary_sparse` named vectors at upsert time). Pages without a summary
+ * surface body-only scores; callers fall back to body-only fusion for those.
+ *
  * Each channel returns up to `limit` hits. A slug is included in the result
- * if it appears in either channel; the missing channel's score is left
- * `undefined` so callers can detect single-channel matches.
+ * if it appears in any channel; missing channel scores stay `undefined` so
+ * callers can distinguish "no match in this channel" from "match with score 0".
  *
  * `restrictToSlugs`, when provided, filters the search server-side to only
  * those slugs (Qdrant `slug IN [...]` filter). Used by `simBatch` when the
@@ -355,42 +506,51 @@ export async function hybridQueryConceptPages(
   // Qdrant 1.13.x sparse-index crash that we've reproduced in the wild.
   const skipSparse = options?.skipSparse ?? false;
 
-  const denseQuery = () =>
+  const queryDense = (using: string) =>
     client.query(MEMORY_V2_COLLECTION, {
       query: dense,
-      using: "dense",
+      using,
       limit,
       with_payload: true,
       filter,
     });
-  const sparseQuery = () =>
+  const querySparse = (using: string) =>
     client.query(MEMORY_V2_COLLECTION, {
       query: sparse,
-      using: "sparse",
+      using,
       limit,
       with_payload: true,
       filter,
     });
 
-  // Run both queries concurrently — they hit independent named vectors.
-  // When sparse is gated off we still resolve a Promise so the destructuring
-  // below stays uniform; the empty `points: []` matches the shape of a
-  // no-hit Qdrant response.
+  // Run all four channels concurrently — they hit independent named vectors.
+  // When sparse is gated off the sparse channels still resolve a Promise so
+  // the destructuring below stays uniform; the empty `points: []` matches
+  // the shape of a no-hit Qdrant response.
   const emptyResult = {
     points: [] as Array<{ payload?: unknown; score?: number }>,
   };
   const runQueries = async () =>
-    Promise.all([denseQuery(), skipSparse ? emptyResult : sparseQuery()]);
+    Promise.all([
+      queryDense("dense"),
+      skipSparse ? emptyResult : querySparse("sparse"),
+      queryDense("summary_dense"),
+      skipSparse ? emptyResult : querySparse("summary_sparse"),
+    ]);
 
   let denseResults;
   let sparseResults;
+  let summaryDenseResults;
+  let summarySparseResults;
   try {
-    [denseResults, sparseResults] = await runQueries();
+    [denseResults, sparseResults, summaryDenseResults, summarySparseResults] =
+      await runQueries();
   } catch (err) {
     if (isCollectionMissing(err)) {
       _collectionReady = false;
       await ensureConceptPageCollection();
-      [denseResults, sparseResults] = await runQueries();
+      [denseResults, sparseResults, summaryDenseResults, summarySparseResults] =
+        await runQueries();
     } else {
       throw err;
     }
@@ -399,21 +559,22 @@ export async function hybridQueryConceptPages(
   // Merge by slug. Missing-side scores stay undefined so the fuser can tell
   // "no match in this channel" apart from "match with score 0".
   const merged = new Map<string, ConceptPageQueryResult>();
-  for (const point of denseResults.points ?? []) {
-    const slug = (point.payload as { slug?: unknown } | null)?.slug;
-    if (typeof slug !== "string") continue;
-    merged.set(slug, { slug, denseScore: point.score ?? 0 });
-  }
-  for (const point of sparseResults.points ?? []) {
-    const slug = (point.payload as { slug?: unknown } | null)?.slug;
-    if (typeof slug !== "string") continue;
-    const existing = merged.get(slug);
-    if (existing) {
-      existing.sparseScore = point.score ?? 0;
-    } else {
-      merged.set(slug, { slug, sparseScore: point.score ?? 0 });
+  const recordHit = (
+    points: Array<{ payload?: unknown; score?: number }> | undefined,
+    set: (entry: ConceptPageQueryResult, score: number) => void,
+  ): void => {
+    for (const point of points ?? []) {
+      const slug = (point.payload as { slug?: unknown } | null)?.slug;
+      if (typeof slug !== "string") continue;
+      const existing = merged.get(slug) ?? { slug };
+      set(existing, point.score ?? 0);
+      merged.set(slug, existing);
     }
-  }
+  };
+  recordHit(denseResults.points, (e, s) => (e.denseScore = s));
+  recordHit(sparseResults.points, (e, s) => (e.sparseScore = s));
+  recordHit(summaryDenseResults.points, (e, s) => (e.summaryDenseScore = s));
+  recordHit(summarySparseResults.points, (e, s) => (e.summarySparseScore = s));
 
   return Array.from(merged.values());
 }

package/src/memory/v2/sim.ts

@@ -120,14 +120,18 @@ export function effectiveWeights(
  *      sparse via the in-process TF-IDF encoder).
  *   2. Run server-side dense + sparse queries against the v2 concept-page
  *      Qdrant collection, restricted to `candidateSlugs` so we don't waste
- *      query bandwidth on unrelated pages.
- *   3. Fuse: per slug, `score = clamp01(dense_weight · denseCosine +
- *      sparse_weight · normalizedSparse)`. Sparse scores are normalized by
- *      the per-batch maximum (so the largest is 1.0); slugs missing from a
- *      channel contribute 0 from that channel.
+ *      query bandwidth on unrelated pages. The query hits four channels per
+ *      page: body dense + body sparse, and (for pages that have a summary
+ *      embedded) summary dense + summary sparse.
+ *   3. Fuse: per slug, score = `max(fused(body), fused(summary))`. Each
+ *      half is `clamp01(dense_weight · denseCosine + sparse_weight ·
+ *      normalizedSparse)` with sparse normalized by the per-batch maximum.
+ *      Pages without a summary embedding fall back to body-only fusion —
+ *      the summary half is undefined and the max collapses to the body
+ *      score.
  *
  * Returns a `Map<slug, score>` containing only the candidate slugs that hit
- * in at least one channel. Slugs in `candidateSlugs` that miss both channels
+ * in at least one channel. Slugs in `candidateSlugs` that miss every channel
  * are absent from the map; callers should treat absence as score = 0 (the
  * activation pipeline does this implicitly when reading back A_o).
  *
@@ -181,20 +185,52 @@ export async function simBatch(
     return new Map();
   }
 
-  const maxSparse = computeMaxSparse(hits);
+  // Compute per-batch sparse maxima independently for the body and summary
+  // channels so each side normalizes against its own scale. Mixing the two
+  // — e.g. dividing every sparse score by the larger of the two maxima —
+  // would punish whichever channel happened to have lower-magnitude scores
+  // even when its hits were the best matches available.
+  const maxBodySparse = computeMaxSparse(hits, (h) => h.sparseScore);
+  const maxSummarySparse = computeMaxSparse(hits, (h) => h.summarySparseScore);
   const { dense_weight: baseDense, sparse_weight: baseSparse } =
     config.memory.v2;
-  const { dense: denseWeight, sparse: sparseWeight } = effectiveWeights(
-    hits,
-    maxSparse,
+  const { dense: bodyDenseWeight, sparse: bodySparseWeight } = effectiveWeights(
+    hits.map((h) => ({ sparseScore: h.sparseScore })),
+    maxBodySparse,
     baseDense,
     baseSparse,
     config,
   );
+  const { dense: summaryDenseWeight, sparse: summarySparseWeight } =
+    effectiveWeights(
+      hits.map((h) => ({ sparseScore: h.summarySparseScore })),
+      maxSummarySparse,
+      baseDense,
+      baseSparse,
+      config,
+    );
 
   const scores = new Map<string, number>();
   for (const hit of hits) {
-    scores.set(hit.slug, fuseHit(hit, maxSparse, denseWeight, sparseWeight));
+    const bodyScore = fuseHalf(
+      hit.denseScore,
+      hit.sparseScore,
+      maxBodySparse,
+      bodyDenseWeight,
+      bodySparseWeight,
+    );
+    const summaryScore = fuseHalf(
+      hit.summaryDenseScore,
+      hit.summarySparseScore,
+      maxSummarySparse,
+      summaryDenseWeight,
+      summarySparseWeight,
+    );
+    // Pages without a summary embedding return undefined for both summary
+    // channels; their `summaryScore` falls back to the body score so the
+    // max collapses cleanly to body-only behavior.
+    const score = Math.max(bodyScore ?? 0, summaryScore ?? bodyScore ?? 0);
+    scores.set(hit.slug, score);
   }
 
   return scores;
@@ -207,36 +243,41 @@ function throwIfAborted(signal: AbortSignal | undefined): void {
 }
 
 /**
- * Per-batch sparse-score maximum used for normalization. Hits missing from
- * the sparse channel contribute 0 (handled by the `undefined` guard).
+ * Per-batch sparse-score maximum used for normalization. The accessor picks
+ * which sparse channel to scan — `sparseScore` for the body channel,
+ * `summarySparseScore` for the summary channel. Hits missing from the
+ * channel contribute 0 (handled by the `undefined` guard).
  */
-function computeMaxSparse(
-  hits: ReadonlyArray<{ sparseScore?: number }>,
+function computeMaxSparse<T>(
+  hits: ReadonlyArray<T>,
+  accessor: (hit: T) => number | undefined,
 ): number {
   let max = 0;
   for (const hit of hits) {
-    if (hit.sparseScore !== undefined && hit.sparseScore > max) {
-      max = hit.sparseScore;
+    const value = accessor(hit);
+    if (value !== undefined && value > max) {
+      max = value;
     }
   }
   return max;
 }
 
 /**
- * Fuse a single hit's dense + sparse scores into a normalized [0, 1] score
+ * Fuse one half of a hit (body or summary) into a normalized [0, 1] score
  * via `clamp01(dense_weight · dense + sparse_weight · sparse/maxSparse)`.
- * Missing-channel scores contribute 0.
+ * Returns `undefined` when neither channel hit — a signal the half had no
+ * match at all, so the caller can fall back to the other half cleanly.
  */
-function fuseHit(
-  hit: { denseScore?: number; sparseScore?: number },
+function fuseHalf(
+  denseScore: number | undefined,
+  sparseScore: number | undefined,
   maxSparse: number,
   denseWeight: number,
   sparseWeight: number,
-): number {
-  const dense = hit.denseScore ?? 0;
+): number | undefined {
+  if (denseScore === undefined && sparseScore === undefined) return undefined;
+  const dense = denseScore ?? 0;
   const sparseNormalized =
-    hit.sparseScore !== undefined && maxSparse > 0
-      ? hit.sparseScore / maxSparse
-      : 0;
+    sparseScore !== undefined && maxSparse > 0 ? sparseScore / maxSparse : 0;
   return clamp01(denseWeight * dense + sparseWeight * sparseNormalized);
 }

package/src/memory/v2/static-context.ts

@@ -18,7 +18,6 @@
 // matching the existing PKB auto-inject pattern.
 
 import type { ChannelId } from "../../channels/types.js";
-import { isAssistantFeatureFlagEnabled } from "../../config/assistant-feature-flags.js";
 import { loadConfig } from "../../config/loader.js";
 import { readPromptFile } from "../../prompts/system-prompt.js";
 import { getWorkspacePromptPath } from "../../util/platform.js";
@@ -36,9 +35,9 @@ const MEMORY_V2_STATIC_BLOCKS: readonly MemoryV2StaticBlock[] = [
 ];
 
 /**
- * Build the v2 static memory block, gated on `memory-v2-enabled` +
- * `config.memory.v2.enabled`. Empty/missing files are skipped; returns
- * `null` when the gate is off or every file is empty.
+ * Build the v2 static memory block, gated on `config.memory.v2.enabled`.
+ * Empty/missing files are skipped; returns `null` when the gate is off or
+ * every file is empty.
  */
 export function readMemoryV2StaticContent(): string | null {
   let config;
@@ -47,10 +46,7 @@ export function readMemoryV2StaticContent(): string | null {
   } catch {
     return null;
   }
-  if (
-    !isAssistantFeatureFlagEnabled("memory-v2-enabled", config) ||
-    !config.memory.v2.enabled
-  ) {
+  if (!config.memory.v2.enabled) {
     return null;
   }
 

package/src/memory/v2/sweep-job.ts

@@ -13,10 +13,10 @@
  * extraction-trigger path. Until then this handler is invoked only by
  * `memory_v2_sweep` rows enqueued explicitly (tests, future CLI).
  *
- * Skipped entirely when the `memory-v2-enabled` feature flag is off, or when
+ * Skipped entirely when `config.memory.v2.enabled` is false, or when
  * `config.memory.v2.sweep_enabled` is false — keeps the sweep dormant in
  * v1-only workspaces and in v2 workspaces that haven't opted in, even if a
- * stale row sits in the queue at flag-flip time.
+ * stale row sits in the queue when v2 is disabled.
  */
 
 import { readFileSync } from "node:fs";
@@ -25,7 +25,6 @@ import { join } from "node:path";
 import { desc, gt } from "drizzle-orm";
 import { z } from "zod";
 
-import { isAssistantFeatureFlagEnabled } from "../../config/assistant-feature-flags.js";
 import type { AssistantConfig } from "../../config/types.js";
 import { getAssistantName } from "../../daemon/identity-helpers.js";
 import {
@@ -104,12 +103,12 @@ export async function memoryV2SweepJob(
   _job: MemoryJob,
   config: AssistantConfig,
 ): Promise<number> {
-  if (!isAssistantFeatureFlagEnabled("memory-v2-enabled", config)) {
-    log.debug("memory-v2-enabled flag off; sweep skipped");
+  if (!config.memory?.v2?.enabled) {
+    log.debug("memory.v2.enabled is false; sweep skipped");
     return 0;
   }
 
-  if (!config.memory?.v2?.sweep_enabled) {
+  if (!config.memory.v2.sweep_enabled) {
     log.debug("memory.v2.sweep_enabled is false; sweep skipped");
     return 0;
   }

package/src/memory/v2/types.ts

@@ -26,10 +26,17 @@ import { z } from "zod";
  * B → A. The full graph is the union of every page's `edges:` list — there
  * is no separate edges-index file. `ref_files` lists paths to attached media
  * (images, audio, etc.).
+ *
+ * `summary` is a 1-4 sentence prose description of the page. When present,
+ * retrieval injects the path + summary instead of the full page so the agent
+ * can decide whether to read the file. Optional because legacy pages predating
+ * the summary field still parse — those fall back to full-page injection and
+ * full-page-only similarity.
  */
 export const ConceptPageFrontmatterSchema = z.object({
   edges: z.array(z.string()).default([]),
   ref_files: z.array(z.string()).default([]),
+  summary: z.string().optional(),
 });
 
 export type ConceptPageFrontmatter = z.infer<