@vellumai/assistant 0.6.6 → 0.7.1

package/src/memory/v2/migration.ts

@@ -0,0 +1,647 @@
+/**
+ * Memory v2 — One-shot v1→v2 migration.
+ *
+ * Gathers v1 graph nodes plus PKB markdown content, clusters them by topic,
+ * synthesizes a concept page per cluster via the configured LLM, promotes
+ * high-significance nodes to `essentials.md` / active follow-ups to
+ * `threads.md` / low-significance episodes to `archive/migrated-<date>.md`,
+ * preserves v1 weighted directional edges as outgoing-edge entries on each
+ * source page's frontmatter, and enqueues `embed_concept_page` jobs for each
+ * new page. A sentinel file at `memory/.v2-state/.migration-complete-v1-to-v2`
+ * gates re-runs — `force: true` is required to overwrite.
+ *
+ * The migration is structured as a sequence of small helpers — `gatherV1State`,
+ * `clusterByTopic`, `synthesizeConceptPage`, `derivePromotions`,
+ * `collapseEdges`, `enqueueEmbeds` — each individually testable so the live
+ * LLM never has to be invoked from the unit suite. `runMemoryV2Migration`
+ * threads them together and writes the sentinel on success.
+ */
+
+import { existsSync, readdirSync, readFileSync } from "node:fs";
+import { mkdir, readFile, writeFile } from "node:fs/promises";
+import { join } from "node:path";
+
+import type { AssistantConfig } from "../../config/types.js";
+import {
+  extractText,
+  getConfiguredProvider,
+  userMessage,
+} from "../../providers/provider-send-message.js";
+import type { Provider } from "../../providers/types.js";
+import { getLogger } from "../../util/logger.js";
+import { type DrizzleDb, getSqliteFrom } from "../db-connection.js";
+import { enqueueMemoryJob } from "../jobs-store.js";
+import { slugify, writePage } from "./page-store.js";
+import type { ConceptPage } from "./types.js";
+
+const log = getLogger("memory-v2-migration");
+
+/** Sentinel file written when the v1→v2 migration completes successfully. */
+export const MIGRATION_SENTINEL_RELATIVE = join(
+  "memory",
+  ".v2-state",
+  ".migration-complete-v1-to-v2",
+);
+
+/**
+ * Result returned by `runMemoryV2Migration`. Counts every meaningful
+ * side-effect so the caller (CLI, IPC route, tests) can surface a useful
+ * summary without having to re-walk the workspace.
+ */
+export interface MigrationResult {
+  pagesCreated: number;
+  edgesWritten: number;
+  essentialsLines: number;
+  threadsLines: number;
+  archiveLines: number;
+  embedsEnqueued: number;
+  sentinelWritten: boolean;
+}
+
+// ---------------------------------------------------------------------------
+// Stage 1 — Gather v1 state
+// ---------------------------------------------------------------------------
+
+/**
+ * v1 source item — either a graph node or a PKB markdown chunk. The shape is
+ * intentionally narrow: each stage only needs `id` (for stable identity),
+ * `text` (for clustering + synthesis), and the structural fields that drive
+ * promotion decisions (`significance`, `type`, `eventDate`).
+ */
+export interface V1Item {
+  id: string;
+  text: string;
+  /** v1 source bucket — drives clustering fallback and promotion decisions. */
+  source: "graph_node" | "pkb_buffer" | "pkb_archive" | "pkb_topic";
+  /** Significance in [0, 1] for graph nodes; 0 for PKB items. */
+  significance: number;
+  /** Memory type for graph nodes; null for PKB items. */
+  type: string | null;
+  /** Epoch ms; null when unknown. */
+  eventDate: number | null;
+  /** Source filename (relative to workspace root) for PKB items; null for graph nodes. */
+  sourcePath: string | null;
+}
+
+/**
+ * v1 weighted directional edge. v2 preserves direction (an edge A→B in v1
+ * becomes an outgoing-edge entry on A's page in v2), so the migration just
+ * has to map node ids to slugs and group by source.
+ */
+export interface V1Edge {
+  sourceNodeId: string;
+  targetNodeId: string;
+}
+
+/**
+ * Read every `memory_graph_nodes` row, every `memory_graph_edges` row, and
+ * every `pkb/` markdown file inside `workspaceDir`. PKB files that don't
+ * exist on disk are silently skipped — a fresh workspace with no PKB content
+ * still migrates cleanly (it just produces no concept pages).
+ */
+export function gatherV1State(
+  database: DrizzleDb,
+  workspaceDir: string,
+): { items: V1Item[]; edges: V1Edge[] } {
+  const items: V1Item[] = [];
+  const raw = getSqliteFrom(database);
+
+  // -- Graph nodes --
+  // Use a raw SELECT so this module stays decoupled from the v1 graph-store
+  // domain layer (which v2 will eventually replace). We only need id /
+  // content / type / significance / event_date — everything else is unused
+  // by the migration.
+  const nodeRows = raw
+    .query<
+      {
+        id: string;
+        content: string;
+        type: string;
+        significance: number;
+        event_date: number | null;
+      },
+      []
+    >(
+      /*sql*/ `SELECT id, content, type, significance, event_date FROM memory_graph_nodes`,
+    )
+    .all();
+  for (const row of nodeRows) {
+    items.push({
+      id: row.id,
+      text: row.content,
+      source: "graph_node",
+      significance: row.significance,
+      type: row.type,
+      eventDate: row.event_date,
+      sourcePath: null,
+    });
+  }
+
+  // -- Graph edges --
+  const edgeRows = raw
+    .query<
+      { source_node_id: string; target_node_id: string },
+      []
+    >(/*sql*/ `SELECT source_node_id, target_node_id FROM memory_graph_edges`)
+    .all();
+  const edges: V1Edge[] = edgeRows.map((row) => ({
+    sourceNodeId: row.source_node_id,
+    targetNodeId: row.target_node_id,
+  }));
+
+  // -- PKB content --
+  const pkbDir = join(workspaceDir, "pkb");
+  if (existsSync(pkbDir)) {
+    items.push(...readPkbItems(pkbDir));
+  }
+
+  return { items, edges };
+}
+
+/**
+ * Walk `pkbDir` once and emit a `V1Item` per markdown file. Sub-buckets:
+ *
+ *   - `buffer.md` (top-level) → `pkb_buffer`
+ *   - `archive/*.md`          → `pkb_archive`
+ *   - everything else `*.md`  → `pkb_topic`
+ *
+ * We deliberately read the full file rather than chunking — the LLM in stage
+ * 3 will summarize, and the upstream PKB chunker is overkill here.
+ */
+function readPkbItems(pkbDir: string): V1Item[] {
+  const items: V1Item[] = [];
+
+  const bufferPath = join(pkbDir, "buffer.md");
+  if (existsSync(bufferPath)) {
+    items.push({
+      id: "pkb:buffer",
+      text: readFileSync(bufferPath, "utf-8"),
+      source: "pkb_buffer",
+      significance: 0,
+      type: null,
+      eventDate: null,
+      sourcePath: "pkb/buffer.md",
+    });
+  }
+
+  const archiveDir = join(pkbDir, "archive");
+  if (existsSync(archiveDir)) {
+    for (const name of readdirSync(archiveDir).sort()) {
+      if (!name.endsWith(".md")) continue;
+      items.push({
+        id: `pkb:archive:${name}`,
+        text: readFileSync(join(archiveDir, name), "utf-8"),
+        source: "pkb_archive",
+        significance: 0,
+        type: null,
+        eventDate: null,
+        sourcePath: join("pkb", "archive", name),
+      });
+    }
+  }
+
+  // Topic files — anything else at the top level besides buffer.md.
+  for (const name of readdirSync(pkbDir).sort()) {
+    if (!name.endsWith(".md")) continue;
+    if (name === "buffer.md") continue;
+    items.push({
+      id: `pkb:topic:${name}`,
+      text: readFileSync(join(pkbDir, name), "utf-8"),
+      source: "pkb_topic",
+      significance: 0,
+      type: null,
+      eventDate: null,
+      sourcePath: join("pkb", name),
+    });
+  }
+
+  return items;
+}
+
+// ---------------------------------------------------------------------------
+// Stage 2 — Cluster by topic
+// ---------------------------------------------------------------------------
+
+/**
+ * One proposed concept page. `slugHint` seeds `slugify` in stage 3 — the
+ * synthesizer is free to refine it but always produces *some* slug.
+ */
+export interface Cluster {
+  slugHint: string;
+  items: V1Item[];
+}
+
+/**
+ * Group v1 items into proposed concept pages.
+ *
+ * The heuristic is intentionally simple — embedding-cluster is on the
+ * "follow-up improvements" list (see plan §11), but for the v1-of-v2 cutover
+ * file-based grouping for PKB plus per-graph-node singletons gives a
+ * reasonable starting set that the LLM can refine in stage 3:
+ *
+ *   - Each `pkb_topic` file becomes its own cluster (slug derived from
+ *     filename — that's literally what topic files were already keyed on).
+ *   - `pkb_buffer` and every `pkb_archive` entry becomes its own cluster
+ *     (these are timeline-style; no obvious topic to merge on).
+ *   - Each graph node is its own cluster (the LLM call in stage 3 already
+ *     synthesizes prose; merging duplicates is consolidation's job).
+ *
+ * The output is deterministic: clusters are emitted in iteration order of
+ * the input items, which matches the SELECT order from `gatherV1State`.
+ */
+export function clusterByTopic(items: V1Item[]): Cluster[] {
+  return items.map((item) => ({
+    slugHint: deriveSlugHint(item),
+    items: [item],
+  }));
+}
+
+/**
+ * Pick a slug seed for a single v1 item. PKB topic files use their filename
+ * (sans `.md`); archive entries use the date stamp; graph nodes use the
+ * first ~6 words of the content. `slugify` enforces ASCII / kebab-case /
+ * length cap downstream so we don't have to.
+ */
+function deriveSlugHint(item: V1Item): string {
+  if (item.source === "pkb_topic" && item.sourcePath) {
+    const base = item.sourcePath.split("/").pop() ?? "";
+    return base.replace(/\.md$/, "");
+  }
+  if (item.source === "pkb_archive" && item.sourcePath) {
+    return `archive-${item.sourcePath.split("/").pop()?.replace(/\.md$/, "")}`;
+  }
+  if (item.source === "pkb_buffer") {
+    return "pkb-buffer";
+  }
+  // graph_node — first few words of the content.
+  return item.text.split(/\s+/).slice(0, 6).join("-");
+}
+
+// ---------------------------------------------------------------------------
+// Stage 3 — Synthesize a concept page per cluster
+// ---------------------------------------------------------------------------
+
+/**
+ * Synthesis prompt — kept here (rather than under `prompts/`) because it's
+ * only used by the migration. The model is asked to emit prose only; the
+ * frontmatter is filled in deterministically by the caller.
+ */
+const SYNTHESIS_SYSTEM_PROMPT = `You are migrating an assistant's memory from a legacy graph + PKB store to a v2 concept-page store. For each cluster of v1 source items, write a single concept page in first-person prose, in the assistant's voice. Do NOT include YAML frontmatter — the caller will add it. Do NOT include a title heading — the slug filename is the title. Keep the body under 5000 characters. Stay grounded in the source items; do not invent facts.`;
+
+/**
+ * Synthesize a single concept page from a `Cluster`. Calls the configured
+ * provider via `memoryV2Migration` LLMCallSite. The body is the model's
+ * response text; the slug is derived from the cluster's `slugHint`.
+ *
+ * `provider` is injected so unit tests can pass a stub. Production callers
+ * pass the result of `getConfiguredProvider("memoryV2Migration")`.
+ *
+ * `identityContext` (typically the assistant's SOUL.md / IDENTITY.md
+ * concatenation) is appended to the system prompt so the synthesized prose
+ * sounds like the assistant rather than a generic narrator.
+ */
+export async function synthesizeConceptPage(
+  cluster: Cluster,
+  identityContext: string | null,
+  provider: Provider,
+): Promise<ConceptPage> {
+  const sourceListing = cluster.items
+    .map((item, i) => {
+      const tag =
+        item.source === "graph_node"
+          ? `node ${item.id} (type=${item.type ?? "?"}, sig=${item.significance.toFixed(2)})`
+          : (item.sourcePath ?? item.source);
+      return `### Source ${i + 1} [${tag}]\n${item.text.trim()}`;
+    })
+    .join("\n\n");
+
+  const systemPrompt = identityContext
+    ? `${SYNTHESIS_SYSTEM_PROMPT}\n\n## Assistant identity\n${identityContext}`
+    : SYNTHESIS_SYSTEM_PROMPT;
+
+  const response = await provider.sendMessage(
+    [
+      userMessage(
+        `Synthesize a single concept page from these v1 sources. Slug hint: \`${cluster.slugHint}\`.\n\n${sourceListing}`,
+      ),
+    ],
+    [],
+    systemPrompt,
+    { config: { callSite: "memoryV2Migration" as const } },
+  );
+  const body = extractText(response);
+
+  return {
+    slug: slugify(cluster.slugHint),
+    frontmatter: { edges: [], ref_files: [] },
+    body: body.endsWith("\n") ? body : `${body}\n`,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Stage 4 — Derive essentials / threads / archive promotions
+// ---------------------------------------------------------------------------
+
+/**
+ * Bucketed promotions written to the prose files in `memory/`. The runner
+ * appends each bucket to its target file (rather than overwriting) so an
+ * operator running `--force` keeps any hand-edits the assistant has made
+ * since the last migration.
+ */
+export interface Promotions {
+  /** Lines for `memory/essentials.md`. */
+  essentials: string[];
+  /** Lines for `memory/threads.md`. */
+  threads: string[];
+  /** Lines for `memory/archive/migrated-<date>.md`. */
+  archive: string[];
+}
+
+/** Cutoff above which a graph node is treated as essential. */
+const ESSENTIALS_SIGNIFICANCE_THRESHOLD = 0.85;
+/** Cutoff below which a graph node is bucketed as low-significance episodic. */
+const ARCHIVE_SIGNIFICANCE_THRESHOLD = 0.3;
+
+/**
+ * Decide which v1 graph nodes get promoted to which prose file. PKB items
+ * are always synthesized into concept pages — they don't get a separate
+ * prose-file promotion.
+ *
+ *   - significance >= 0.85         → `essentials.md`
+ *   - type == "prospective"        → `threads.md`
+ *   - significance <= 0.3          → archive
+ *   - everything else              → no promotion (only the synthesized page)
+ *
+ * The thresholds are intentionally chunky — fine-tuning happens during
+ * consolidation, not here.
+ */
+export function derivePromotions(items: V1Item[]): Promotions {
+  const essentials: string[] = [];
+  const threads: string[] = [];
+  const archive: string[] = [];
+
+  for (const item of items) {
+    if (item.source !== "graph_node") continue;
+
+    const summary = formatPromotionLine(item);
+    if (item.significance >= ESSENTIALS_SIGNIFICANCE_THRESHOLD) {
+      essentials.push(summary);
+      continue;
+    }
+    if (item.type === "prospective") {
+      threads.push(summary);
+      continue;
+    }
+    if (item.significance <= ARCHIVE_SIGNIFICANCE_THRESHOLD) {
+      archive.push(summary);
+    }
+  }
+
+  return { essentials, threads, archive };
+}
+
+/** Format a single graph node as a one-line bullet for the prose files. */
+function formatPromotionLine(item: V1Item): string {
+  // Strip newlines so each item becomes a single bullet.
+  const text = item.text.replace(/\s+/g, " ").trim();
+  return `- ${text}`;
+}
+
+// ---------------------------------------------------------------------------
+// Stage 5 — Map v1 edges to per-page outgoing edges
+// ---------------------------------------------------------------------------
+
+/**
+ * Map every v1 graph-node id to a v2 concept-page slug and group v1 edges by
+ * source slug. Edges with either endpoint missing from `slugMap` are dropped
+ * silently — those endpoints didn't survive synthesis (e.g. their cluster
+ * produced no usable page). Self-loops are dropped; duplicate `(source, target)`
+ * pairs collapse via the `Set<string>`.
+ *
+ * The returned map keys are source slugs and the values are sets of target
+ * slugs — i.e. each entry is a page's outgoing-edge list. The runner writes
+ * these into the source page's frontmatter.
+ */
+export function collapseEdges(
+  v1Edges: V1Edge[],
+  slugMap: Map<string, string>,
+): Map<string, Set<string>> {
+  const outgoing = new Map<string, Set<string>>();
+  for (const edge of v1Edges) {
+    const from = slugMap.get(edge.sourceNodeId);
+    const to = slugMap.get(edge.targetNodeId);
+    if (!from || !to) continue;
+    if (from === to) continue;
+    let targets = outgoing.get(from);
+    if (!targets) {
+      targets = new Set<string>();
+      outgoing.set(from, targets);
+    }
+    targets.add(to);
+  }
+  return outgoing;
+}
+
+// ---------------------------------------------------------------------------
+// Stage 6 — Fan out embed jobs
+// ---------------------------------------------------------------------------
+
+/**
+ * Enqueue an `embed_concept_page` job for each newly-written slug. The job
+ * itself lands in PR 13 — we just stage the queue here, so the embeddings
+ * are ready by the time activation needs them.
+ */
+export function enqueueEmbeds(slugs: string[]): number {
+  for (const slug of slugs) {
+    enqueueMemoryJob("embed_concept_page", { slug });
+  }
+  return slugs.length;
+}
+
+// ---------------------------------------------------------------------------
+// Top-level migration runner
+// ---------------------------------------------------------------------------
+
+export interface RunMemoryV2MigrationParams {
+  workspaceDir: string;
+  database: DrizzleDb;
+  /** Overwrite existing v2 state when the sentinel is already present. */
+  force?: boolean;
+  /** Identity context appended to the synthesis system prompt. */
+  identityContext?: string | null;
+  /** Caller-supplied LLM provider for synthesis. Defaults to the configured provider. */
+  provider?: Provider;
+  /** Caller-supplied config; reserved for future tunables (e.g. cluster knobs). */
+  config?: AssistantConfig;
+}
+
+/**
+ * Run the full v1→v2 migration. Returns a `MigrationResult` summarizing
+ * every side-effect for the CLI / IPC caller.
+ *
+ * Re-runs are gated by the sentinel file
+ * `memory/.v2-state/.migration-complete-v1-to-v2`. Without `force: true`, a
+ * second invocation throws `MigrationAlreadyAppliedError` without mutating
+ * anything; with `force: true`, the migration overwrites pages and re-appends
+ * to the prose files.
+ */
+export class MigrationAlreadyAppliedError extends Error {
+  constructor() {
+    super(
+      "Memory v2 migration sentinel exists; pass force: true to re-run. Running without --force preserves the existing v2 state.",
+    );
+    this.name = "MigrationAlreadyAppliedError";
+  }
+}
+
+export async function runMemoryV2Migration(
+  params: RunMemoryV2MigrationParams,
+): Promise<MigrationResult> {
+  const {
+    workspaceDir,
+    database,
+    force = false,
+    identityContext = null,
+  } = params;
+
+  const sentinelPath = join(workspaceDir, MIGRATION_SENTINEL_RELATIVE);
+  if (existsSync(sentinelPath) && !force) {
+    throw new MigrationAlreadyAppliedError();
+  }
+
+  const { items, edges: v1Edges } = gatherV1State(database, workspaceDir);
+  log.info(
+    { itemCount: items.length, edgeCount: v1Edges.length },
+    "Gathered v1 state for memory v2 migration",
+  );
+
+  const clusters = clusterByTopic(items);
+
+  const provider =
+    params.provider ?? (await getConfiguredProvider("memoryV2Migration"));
+  if (!provider) {
+    throw new Error(
+      "memoryV2Migration provider unavailable — configure llm.callSites.memoryV2Migration or llm.default before re-running.",
+    );
+  }
+
+  // PKB items never become edge endpoints, so the slug map only needs the
+  // graph-node side of the cluster.
+  const slugMap = new Map<string, string>();
+  const pages: ConceptPage[] = [];
+  const usedSlugs = new Set<string>();
+
+  for (const cluster of clusters) {
+    const page = await synthesizeConceptPage(
+      cluster,
+      identityContext,
+      provider,
+    );
+    // Disambiguate slug collisions deterministically: append `-2`, `-3`, …
+    // until we find an unused slug. Keeps synthesizeConceptPage stateless
+    // while still tolerating two clusters that hash to the same hint.
+    let finalSlug = page.slug;
+    let suffix = 2;
+    while (usedSlugs.has(finalSlug)) {
+      finalSlug = `${page.slug}-${suffix++}`;
+    }
+    usedSlugs.add(finalSlug);
+    const finalized: ConceptPage = { ...page, slug: finalSlug };
+    pages.push(finalized);
+
+    for (const item of cluster.items) {
+      if (item.source === "graph_node") slugMap.set(item.id, finalSlug);
+    }
+  }
+
+  // Resolve outgoing edges per source slug, then attach them to each page's
+  // frontmatter before writing. The page is the source of truth for its own
+  // outgoing edges — there is no separate edges-index file.
+  const outgoingBySource = collapseEdges(v1Edges, slugMap);
+  const finalizedPages: ConceptPage[] = pages.map((page) => {
+    const targets = outgoingBySource.get(page.slug);
+    if (!targets || targets.size === 0) return page;
+    return {
+      ...page,
+      frontmatter: {
+        ...page.frontmatter,
+        edges: [...targets].sort(),
+      },
+    };
+  });
+
+  // Page writes hit different filenames so they're safe to fan out.
+  await Promise.all(
+    finalizedPages.map((page) => writePage(workspaceDir, page)),
+  );
+
+  const promotions = derivePromotions(items);
+  await appendPromotions(workspaceDir, promotions);
+
+  const embedsEnqueued = enqueueEmbeds(finalizedPages.map((p) => p.slug));
+
+  await writeSentinel(workspaceDir);
+
+  let edgesWritten = 0;
+  for (const targets of outgoingBySource.values()) {
+    edgesWritten += targets.size;
+  }
+
+  return {
+    pagesCreated: finalizedPages.length,
+    edgesWritten,
+    essentialsLines: promotions.essentials.length,
+    threadsLines: promotions.threads.length,
+    archiveLines: promotions.archive.length,
+    embedsEnqueued,
+    sentinelWritten: true,
+  };
+}
+
+/**
+ * Append each promotion bucket to its target file. Files are created if
+ * absent — the workspace migration in PR 3 (`060-memory-v2-init`) seeds
+ * empty placeholders, so this is mostly belt-and-suspenders.
+ */
+async function appendPromotions(
+  workspaceDir: string,
+  promotions: Promotions,
+): Promise<void> {
+  const memoryDir = join(workspaceDir, "memory");
+  await mkdir(memoryDir, { recursive: true });
+  await mkdir(join(memoryDir, "archive"), { recursive: true });
+
+  if (promotions.essentials.length > 0) {
+    await appendLines(join(memoryDir, "essentials.md"), promotions.essentials);
+  }
+  if (promotions.threads.length > 0) {
+    await appendLines(join(memoryDir, "threads.md"), promotions.threads);
+  }
+  if (promotions.archive.length > 0) {
+    const today = new Date().toISOString().slice(0, 10);
+    await appendLines(
+      join(memoryDir, "archive", `migrated-${today}.md`),
+      promotions.archive,
+    );
+  }
+}
+
+/** Append `lines` to `path`, creating it (with a trailing newline) if absent. */
+async function appendLines(path: string, lines: string[]): Promise<void> {
+  let existing = "";
+  try {
+    existing = await readFile(path, "utf-8");
+  } catch (err) {
+    if ((err as NodeJS.ErrnoException).code !== "ENOENT") throw err;
+  }
+  const trailing = existing.length === 0 || existing.endsWith("\n") ? "" : "\n";
+  const next = `${existing}${trailing}${lines.join("\n")}\n`;
+  await writeFile(path, next, "utf-8");
+}
+
+/** Write the migration sentinel. The body is metadata for human inspection. */
+async function writeSentinel(workspaceDir: string): Promise<void> {
+  const sentinelPath = join(workspaceDir, MIGRATION_SENTINEL_RELATIVE);
+  await mkdir(join(workspaceDir, "memory", ".v2-state"), { recursive: true });
+  await writeFile(sentinelPath, `${new Date().toISOString()}\n`, "utf-8");
+}

package/src/memory/v2/now-text.ts

@@ -0,0 +1,37 @@
+// ---------------------------------------------------------------------------
+// Memory v2 — Shared "NOW" context loader
+// ---------------------------------------------------------------------------
+//
+// The activation formula's `c_now` term needs a snapshot of the prose meta
+// files (`essentials.md`, `threads.md`, `recent.md`).
+// Missing or unreadable files are treated as empty so a fresh workspace
+// (no consolidation has run yet) still reaches the v2 injector with a
+// well-defined `nowText`.
+
+import { readFile } from "node:fs/promises";
+import { join } from "node:path";
+
+const NOW_FILES = ["essentials.md", "threads.md", "recent.md"] as const;
+
+/**
+ * Read `memory/{essentials,threads,recent}.md` and concatenate the trimmed
+ * non-empty contents.
+ *
+ * Returns an empty string when none of the files exist or all are empty.
+ */
+export async function loadNowText(workspaceDir: string): Promise<string> {
+  const reads = await Promise.all(
+    NOW_FILES.map(async (filename) => {
+      try {
+        const text = await readFile(
+          join(workspaceDir, "memory", filename),
+          "utf-8",
+        );
+        return text.trim();
+      } catch {
+        return "";
+      }
+    }),
+  );
+  return reads.filter((part) => part.length > 0).join("\n\n");
+}