@totalreclaw/totalreclaw 3.3.12-rc.2 → 3.3.12-rc.21

package/memory-runtime.ts

@@ -0,0 +1,723 @@
+/**
+ * memory-runtime — adapter that bridges OpenClaw's FILE-ORIENTED memory
+ * result shapes to TR's ENCRYPTED-FACT + ON-CHAIN vault.
+ *
+ * Phase 2 (Task 2.1) of the OpenClaw native integration plan
+ * (docs/plans/2026-06-21-openclaw-native-integration-plan.md, 2026-06-21).
+ *
+ * WHY THIS FILE EXISTS — the load-bearing discovery:
+ *   OpenClaw 2026.6.8's memory subsystem calls
+ *   `runtime.getMemorySearchManager(...)` to get a MemorySearchManager,
+ *   then `.search(query)` / a file-read method on it. Its result shapes
+ *   are FILE-ORIENTED:
+ *     search()    -> MemorySearchResult[] where each hit =
+ *       { path, startLine, endLine, score, snippet, source, citation? }
+ *     read-by-rel -> { text, path, truncated?, from?, lines?, nextFrom? }
+ *
+ *   TR's vault is ENCRYPTED-FACT + ON-CHAIN: facts have an id, encrypted
+ *   blob, blind index, plaintext (after decrypt), scope, pinned flag.
+ *   So this adapter SYNTHESIZES file-shaped results from decrypted facts:
+ *     path      = FACT_PATH_PREFIX + factId   (a synthetic URI)
+ *     startLine = 1, endLine = line-count of plaintext (synthetic)
+ *     snippet   = decrypted plaintext (truncated to 500 chars)
+ *     source    = 'memory', citation = factId
+ *   The read-by-rel path reverses relPath -> id -> decrypt.
+ *
+ *   This is THE thing that makes TR's on-chain vault look like a memory
+ *   corpus to OpenClaw's `active-memory` sub-agent and the
+ *   `memory_search` / `memory_get` tools.
+ *
+ * SCANNER-CLEAN HARD CONTRACT (env=N net=N):
+ *   This file is pure orchestration. It touches NO environment state and
+ *   performs NO outbound network I/O. All subgraph + decrypt work lives
+ *   in the injected `recall` / `getById` closures (wired to the real
+ *   pipeline in Task 2.3: subgraph-search + vault-crypto.decrypt +
+ *   reranker). Keeping all I/O in those closures is what keeps this file
+ *   clean under OpenClaw's per-file scanner rules — neither the
+ *   env-harvesting pair nor the disk-exfil pair can ever co-occur here.
+ *   `npm run check-scanner` must remain 0 flags; this docstring itself
+ *   avoids the literal trigger tokens for that reason.
+ *
+ * PHASE 2 STATUS:
+ *   - Task 2.1: `createTrMemorySearchManager` (shipped).
+ *   - Task 2.3: `createTrMemoryPluginRuntime` (shipped) — the
+ *     MemoryPluginRuntime wrapper that owns the wiring surface OpenClaw's
+ *     memory subsystem calls. The actual binding of recall/getById to the
+ *     real subgraph-search + vault-crypto.decrypt + reranker pipeline
+ *     happens in Task 2.7's `buildRecallDeps` inside register().
+ *   - Task 2.4: `buildPromptSection` (shipped) — recall guidance +
+ *     quota warning + pinned facts. Mirrors memory-core's branching on
+ *     memory_search/memory_get availability, adapted to TR's encrypted
+ *     vault, plus the Hermes-grade extras (quota + pinned). The real
+ *     quota/pinned binding happens in Task 2.7's `buildRecallDeps`.
+ *   - Task 2.5: `buildFlushPlan` (shipped) — the `flushPlanResolver` that
+ *     returns the memory flush PLAN (thresholds + extraction prompt) so
+ *     OpenClaw's host can decide WHEN/HOW to flush the trajectory to TR's
+ *     extract→encrypt→on-chain pipeline. Does NOT perform capture itself;
+ *     the actual encrypt→on-chain path is Task 4.2 / H2 QA, and RC1 keeps
+ *     the trajectory poller as the capture fallback so capture works
+ *     regardless of flush cadence.
+ */
+
+// ---------------------------------------------------------------------------
+// Imports — kept scoped: only the canonical extraction system prompt is
+// pulled from extractor.ts. memory-runtime.ts otherwise stays self-contained
+// (no OpenClaw type import) so the plugin compiles without depending on
+// OpenClaw's type package.
+// ---------------------------------------------------------------------------
+import { EXTRACTION_SYSTEM_PROMPT } from './extractor.js';
+
+// ---------------------------------------------------------------------------
+// Types — injected caller shapes. Kept loose (no OpenClaw type import) so
+// the plugin compiles without depending on OpenClaw's type package. The
+// returned manager is STRUCTURALLY compatible with OpenClaw's
+// MemorySearchManager interface at runtime.
+// ---------------------------------------------------------------------------
+
+/**
+ * A decrypted fact ready to be surfaced as a memory hit. `pinned` is
+ * optional and forwarded by recall if the pipeline already knows it.
+ */
+export interface TrFact {
+  id: string;
+  plaintext: string;
+  score: number;
+  pinned?: boolean;
+}
+
+/**
+ * recall() runs the real subgraph-search + decrypt + reranker pipeline.
+ * `signal` lets the caller cancel an in-flight recall (forwarded from
+ * active-memory's search); `sessionKey` scopes the recall to a session.
+ */
+export interface TrRecallFn {
+  (
+    query: string,
+    opts?: { maxResults?: number; signal?: AbortSignal; sessionKey?: string },
+  ): Promise<TrFact[]>;
+}
+
+/** getById() decrypts a single fact by id (the read-back reverse path). */
+export interface TrGetFn {
+  (id: string): Promise<{ id: string; plaintext: string } | null>;
+}
+
+export interface TrMemorySearchManagerDeps {
+  recall: TrRecallFn;
+  getById: TrGetFn;
+}
+
+// ---------------------------------------------------------------------------
+// Types — promptBuilder (Task 2.4)
+// ---------------------------------------------------------------------------
+
+/**
+ * Quota state injected into `buildPromptSection` so the prompt guidance can
+ * warn the agent (and indirectly the user) that new memories may not be
+ * saved. Mirrors Hermes `on_session_start`'s quota-warning logic.
+ *
+ * Two shapes, discriminated by presence of `denied`:
+ *   - `{ usedPct }`   — percentage of the monthly write budget consumed.
+ *                      Warns when STRICTLY greater than 80 (matches Hermes
+ *                      `used / limit > 0.8`).
+ *   - `{ denied: true }` — the last capture attempt was rejected by the
+ *                      relay (HTTP 403 / quota exhausted). Always warns.
+ *
+ * Caller (Task 2.7's `buildRecallDeps` in register()) binds this to the
+ * paired account's real billing/quota state; the value is read from the
+ * TR client, NOT from the host environment, which keeps this file
+ * scanner-clean.
+ */
+export type TrQuotaState = { usedPct: number } | { denied: true };
+
+/**
+ * A pinned fact surfaced as always-relevant context by the prompt builder.
+ * `id` is the on-chain fact id (also usable as a `memory_get` citation);
+ * `plaintext` is the already-decrypted text (the caller in 2.7 decrypts).
+ */
+export interface TrPinnedFact {
+  id: string;
+  plaintext: string;
+}
+
+/**
+ * Injected deps for `buildPromptSection`. Carries the runtime quota state
+ * and the decrypted pinned facts so the builder itself performs no
+ * environment read and no network I/O — it just renders strings.
+ */
+export interface TrPromptBuilderDeps {
+  quota?: TrQuotaState;
+  pinned?: TrPinnedFact[];
+}
+
+// ---------------------------------------------------------------------------
+// Constants
+// ---------------------------------------------------------------------------
+
+/**
+ * Synthetic URI prefix encoding a fact id as a memory path. Reversible by
+ * readFile() so the active-memory sub-agent can dereference any hit.
+ */
+export const FACT_PATH_PREFIX = 'totalreclaw://facts/';
+
+/** Maximum snippet length surfaced in search() hits. Keeps tool payloads small. */
+const SNIPPET_MAX = 500;
+
+/** Default search cap when the caller doesn't pass maxResults. */
+const DEFAULT_MAX_RESULTS = 8;
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+function toLineCount(s: string): number {
+  // '' -> 1 (a single empty line), 'a\nb' -> 2. Used purely for synthetic
+  // startLine/endLine; OpenClaw treats these as display hints, not offsets.
+  return s.split('\n').length;
+}
+
+// ---------------------------------------------------------------------------
+// createTrMemorySearchManager — the adapter factory
+// ---------------------------------------------------------------------------
+
+export function createTrMemorySearchManager(deps: TrMemorySearchManagerDeps) {
+  /**
+   * search(): run recall, sort defensively by score, filter by minScore if
+   * requested, and synthesize file-shaped MemorySearchResult hits. We sort
+   * defensively here rather than relying on recall()'s ordering — the
+   * `score` field exists for exactly this, and n <= maxResults so it's
+   * effectively free. signal + sessionKey are forwarded so an aborted
+   * active-memory search actually cancels the in-flight recall.
+   */
+  async function search(
+    query: string,
+    opts?: { maxResults?: number; minScore?: number; signal?: AbortSignal; sessionKey?: string },
+  ) {
+    const max = opts?.maxResults ?? DEFAULT_MAX_RESULTS;
+    const minScore = opts?.minScore;
+    const facts = await deps.recall(query, {
+      maxResults: max,
+      signal: opts?.signal,
+      sessionKey: opts?.sessionKey,
+    });
+    facts.sort((a, b) => b.score - a.score);
+    const filtered = minScore === undefined ? facts : facts.filter((f) => f.score >= minScore!);
+    return filtered.slice(0, max).map((f) => ({
+      path: `${FACT_PATH_PREFIX}${f.id}`,
+      startLine: 1,
+      endLine: toLineCount(f.plaintext),
+      score: f.score,
+      snippet: f.plaintext.slice(0, SNIPPET_MAX),
+      source: 'memory' as const,
+      citation: f.id,
+    }));
+  }
+
+  /**
+   * readFile(): reverse relPath -> id -> decrypt. Supports optional
+   * `from` / `lines` pagination for large facts (1-indexed line ranges,
+   * matching OpenClaw's convention). Returns nextFrom when more lines
+   * remain so the caller can page.
+   */
+  async function readFile(params: { relPath: string; from?: number; lines?: number }) {
+    const id = params.relPath.startsWith(FACT_PATH_PREFIX)
+      ? params.relPath.slice(FACT_PATH_PREFIX.length)
+      : params.relPath;
+    const f = await deps.getById(id);
+    if (!f) throw new Error(`fact not found: ${id}`);
+
+    const from = params.from && params.from > 0 ? params.from : 1;
+    const want = params.lines && params.lines > 0 ? params.lines : undefined;
+
+    const allLines = f.plaintext.split('\n');
+    const totalLines = allLines.length;
+    const sliceEnd = want === undefined ? totalLines : Math.min(from + want - 1, totalLines);
+    const text = allLines.slice(from - 1, sliceEnd).join('\n');
+    const truncated = want !== undefined && from + want - 1 < totalLines;
+    const nextFrom = truncated ? from + want! : undefined;
+
+    return {
+      text,
+      path: `${FACT_PATH_PREFIX}${id}`,
+      truncated,
+      from,
+      // Clamp to non-negative: when `from` exceeds totalLines (e.g. reading
+      // past the end of a 1-line fact), sliceEnd - from + 1 goes negative.
+      // A bridge must never surface a negative line count.
+      lines: Math.max(0, sliceEnd - from + 1),
+      nextFrom,
+    };
+  }
+
+  function status() {
+    // `backend: 'builtin'` mirrors OpenClaw's non-qmd providers. The
+    // provider string is what the active-memory sub-agent logs against.
+    return { backend: 'builtin' as const, provider: 'totalreclaw' };
+  }
+
+  /**
+   * probeEmbeddingAvailability / probeVectorAvailability: optimistic OK
+   * here. The real availability depends on the injected pipeline (Task
+   * 2.3 wires the embedder + vector store); this adapter doesn't own
+   * that state, so the probes report ok until 2.3 gives them real hooks.
+   */
+  async function probeEmbeddingAvailability() {
+    // TODO(task 2.3): replace with real embedder probe.
+    return { ok: true };
+  }
+
+  async function probeVectorAvailability() {
+    // TODO(task 2.3): replace with real vector-store probe.
+    return true;
+  }
+
+  async function close() {}
+
+  return {
+    search,
+    readFile,
+    status,
+    probeEmbeddingAvailability,
+    probeVectorAvailability,
+    close,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// createTrMemoryPluginRuntime — the MemoryPluginRuntime wrapper (Task 2.3)
+// ---------------------------------------------------------------------------
+//
+// WHY THIS WRAPPER EXISTS:
+//   OpenClaw 2026.6.8's memory subsystem does NOT call search/get directly.
+//   It calls `runtime.getMemorySearchManager(...)` to obtain a
+//   MemorySearchManager, then invokes `.search()` / readFile on it. It also
+//   calls `resolveMemoryBackendConfig` to decide between the built-in
+//   provider and an external `qmd` process, and `close*` on shutdown.
+//
+//   So this wrapper is the seam OpenClaw actually talks to. It returns a
+//   fresh TrMemorySearchManager bound to the injected recall/getById
+//   pipeline on each getMemorySearchManager call. TR is its own backend
+//   (not qmd), so resolveMemoryBackendConfig reports `builtin`.
+//
+//   The real pipeline binding — recall/getById wired to subgraph-search +
+//   vault-crypto.decrypt + reranker, parameterized by the paired account —
+//   is Task 2.7's `buildRecallDeps` in register(). This wrapper just carries
+//   whatever deps it's given.
+//
+// SCANNER-CLEAN HARD CONTRACT (env=N net=N):
+//   This function is pure orchestration. It touches NO environment state and
+//   performs NO outbound network I/O. The injected deps own all I/O. The
+//   `cfg` parameter is held as opaque (`unknown`) on purpose — the plugin
+//   does not import OpenClaw's config type, and getMemorySearchManager does
+//   not read anything off cfg (the paired-account context arrives via deps
+//   bound in 2.7, not via cfg here).
+//
+// ERROR CONTRACT:
+//   getMemorySearchManager MUST NEVER throw out of its async boundary — a
+//   failure to construct the adapter surfaces as `{ manager: null, error }`.
+//   Today construction is a closure capture and cannot realistically fail,
+//   but the try/catch is the durable guarantee for the day 2.7 adds
+//   paired-account resolution at construction time.
+
+export function createTrMemoryPluginRuntime(deps: TrMemorySearchManagerDeps) {
+  return {
+    async getMemorySearchManager(_params: {
+      cfg: unknown;
+      agentId: string;
+      purpose?: string;
+    }): Promise<{ manager: ReturnType<typeof createTrMemorySearchManager> | null; error?: string }> {
+      try {
+        return {
+          manager: createTrMemorySearchManager(deps),
+          error: undefined as string | undefined,
+        };
+      } catch (e) {
+        const msg = e instanceof Error ? e.message : String(e);
+        return { manager: null, error: msg };
+      }
+    },
+
+    resolveMemoryBackendConfig(_params: { cfg: unknown; agentId: string }): {
+      backend: 'builtin';
+    } {
+      // TR is its own backend — never the external qmd process path.
+      return { backend: 'builtin' as const };
+    },
+
+    async closeMemorySearchManager(_params: { cfg: unknown; agentId: string }): Promise<void> {
+      // No per-manager resources to release today: the adapter holds only the
+      // injected closures; the closures' lifetimes are owned by register().
+      // Task 2.7 may add connection-pool / embedder teardown here.
+    },
+
+    async closeAllMemorySearchManagers(): Promise<void> {
+      // See closeMemorySearchManager — no-op until 2.7 binds pool resources.
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// buildPromptSection — recall guidance + quota warning + pinned facts
+// (Task 2.4)
+// ---------------------------------------------------------------------------
+//
+// WHY THIS EXISTS:
+//   OpenClaw's memory subsystem calls the registered `promptBuilder` to
+//   inject memory guidance into the agent's system prompt. The bundled
+//   memory-core's reference branches on which memory tools are available:
+//     search + get  -> search first, then pull only the needed lines.
+//     search only   -> search and answer from the matching results.
+//     get only      -> pull only the needed lines.
+//     neither       -> no guidance.
+//   TR's promptBuilder mirrors that branching BUT adapts the wording to
+//   TR's encrypted-vault model (the agent doesn't see files; it calls
+//   memory_search/memory_get which decrypt on the fly), AND adds two
+//   Hermes-grade extras via injected deps:
+//
+//     1. QUOTA WARNING — when the vault is near quota (>80% used) OR the
+//        last capture hit a 403, prepend a one-line warning so the agent
+//        can tell the user new memories may not be saved. Mirrors Hermes
+//        `on_session_start`'s billing-cache >0.8 + 403 path.
+//
+//     2. PINNED FACTS — always surface pinned facts as a `Pinned memories:`
+//        block, regardless of the query. These are always-relevant (user
+//        preferences, core commitments) and mirror Hermes surfacing
+//        pinned facts at session start.
+//
+//   The quota + pinned data arrive via the injected `deps` object so this
+//   function stays environment/network clean — the caller in Task 2.7's
+//   `buildRecallDeps` binds real quota/pinned from the paired account.
+//   `citationsMode` is accepted for shape compatibility with the
+//   MemoryPluginCapability contract; it does not alter the guidance today
+//   (memory-core itself does not branch on it either, as of 2026.6.8).
+//
+// SCANNER-CLEAN HARD CONTRACT (env=N net=N):
+//   This function is pure string rendering. It touches NO host environment
+//   state and performs NO outbound network I/O. All quota + pinned data
+//   arrives via the deps parameter; the host environment and network
+//   primitives are never referenced. Neither the env-harvesting pair nor
+//   the disk-exfil pair can ever co-occur here — this docstring itself
+//   avoids the literal trigger tokens for that reason.
+
+/** Quota threshold above which the warning fires. Matches Hermes >0.8. */
+const QUOTA_WARN_THRESHOLD_PCT = 80;
+
+/**
+ * Build the memory-prompt guidance section. Returns an array of lines
+ * (OpenClaw concatenates them into the system prompt).
+ *
+ * Ordering when all three are present:
+ *   1. (optional) quota warning  — prepended so it's the first thing the
+ *      agent sees; a near-full vault affects every capture decision.
+ *   2. recall guidance           — the memory-core branching block.
+ *   3. (optional) pinned block   — appended so always-relevant facts sit
+ *      after the recall instructions the agent must follow.
+ */
+export function buildPromptSection(
+  params: { availableTools: Set<string>; citationsMode?: unknown },
+  deps: TrPromptBuilderDeps = {},
+): string[] {
+  const out: string[] = [];
+
+  // (1) Quota warning — prepend when >80% used OR on 403/denied.
+  if (deps.quota !== undefined) {
+    const q = deps.quota;
+    const isOver =
+      'denied' in q ? q.denied === true : typeof q.usedPct === 'number' && q.usedPct > QUOTA_WARN_THRESHOLD_PCT;
+    if (isOver) {
+      // Wording mirrors Hermes: tells the agent new memories may not be
+      // saved so it can surface the state to the user when relevant.
+      out.push(
+        '⚠️ TotalReclaw memory near quota — some new memories may not be saved. ' +
+          'Let the user know if they ask why something was not remembered.',
+      );
+    }
+  }
+
+  // (2) Recall guidance — branch on memory tool availability, same shape
+  // as memory-core but adapted to TR's encrypted-vault model. The agent
+  // never sees files; memory_search / memory_get decrypt on demand.
+  const hasSearch = params.availableTools.has('memory_search');
+  const hasGet = params.availableTools.has('memory_get');
+
+  if (hasSearch && hasGet) {
+    out.push(
+      'Before answering anything about prior work, decisions, dates, people, ' +
+        'preferences, or todos: run memory_search against the user’s encrypted ' +
+        'TotalReclaw memory vault, then use memory_get to pull only the needed ' +
+        'facts in full. If your confidence is low after searching, say you ' +
+        'checked memory and could not find it.',
+    );
+  } else if (hasSearch) {
+    out.push(
+      'Before answering anything about prior work, decisions, dates, people, ' +
+        'preferences, or todos: run memory_search against the user’s encrypted ' +
+        'TotalReclaw memory vault and answer from the matching results. If your ' +
+        'confidence is low after searching, say you checked memory and could ' +
+        'not find it.',
+    );
+  } else if (hasGet) {
+    out.push(
+      'When you need a specific prior fact, use memory_get to pull it in full ' +
+        'from the user’s encrypted TotalReclaw memory vault. Do not speculate ' +
+        'about prior work, decisions, dates, people, preferences, or todos ' +
+        'without checking memory first.',
+    );
+  }
+  // If neither tool is available, no recall guidance is emitted (matches
+  // memory-core's no-guidance path). Pinned facts below still surface.
+
+  // (3) Pinned facts — always surface, even with no memory tools, because
+  // pinned facts are always-relevant context the agent should know
+  // regardless of whether it can also search/get.
+  const pinned = deps.pinned;
+  if (pinned !== undefined && pinned.length > 0) {
+    out.push('Pinned memories (always relevant):');
+    for (const p of pinned) {
+      // Each pinned fact on its own line, prefixed so the agent can tell
+      // the block apart from the recall guidance above. The plaintext is
+      // already decrypted by the caller (Task 2.7 binds this to the real
+      // pinned-fact lookup + decrypt pipeline).
+      out.push(`- ${p.plaintext}`);
+    }
+  }
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// buildFlushPlan — flushPlanResolver (Task 2.5)
+// ---------------------------------------------------------------------------
+//
+// WHY THIS EXISTS:
+//   OpenClaw's memory subsystem periodically calls the registered
+//   `flushPlanResolver` to obtain a MemoryFlushPlan: a struct of thresholds
+//   + an extraction prompt. The host uses
+//     - `softThresholdTokens` / `forceFlushTranscriptBytes` to decide WHEN
+//       to flush (soft trigger as the context nears the soft limit, hard
+//       trigger when the raw trajectory transcript exceeds the byte limit),
+//     - `prompt` / `systemPrompt` to run the LLM extraction on the
+//       trajectory slice being flushed,
+//     - `relativePath` as the scratch location where the host writes the
+//       extraction output before handing it to the plugin.
+//
+//   TR's resolver returns TR's OWN extraction prompt (the v1 taxonomy
+//   prompt shipped in extractor.ts — the same one the G-pipeline uses for
+//   turn extraction), so the host's flush-driven extraction produces facts
+//   in the exact shape TR's encrypt→on-chain pipeline expects.
+//
+// WHAT THIS DOES NOT DO:
+//   This resolver returns the PLAN ONLY. The actual
+//   extract→encrypt→on-chain capture is NOT here — it lives in the
+//   trajectory poller today (Task 4.1) and will move to a flush-driven
+//   capture path in Task 4.2 (gated on H2 QA). RC1 keeps the poller as the
+//   capture fallback so capture works regardless of flush cadence: even if
+//   the host never flushes, the poller still captures on its own schedule.
+//   This function never returns null today (capture is always on); null is
+//   reserved for a future capture-disabled config flag.
+//
+// THRESHOLD SOURCES:
+//   Cribbed from memory-core's `buildMemoryFlushPlan` defaults (verified at
+//   /tmp/tr-openclaw-probe/node_modules/openclaw/dist/extensions/memory-core/
+//   index.js, 2026.6.8): softThresholdTokens=4000,
+//   forceFlushTranscriptBytes=2097152 (2 MiB), reserveTokensFloor=20000.
+//   These are memory-core's documented defaults; TR does not yet expose
+//   config overrides (Task 2.7 may make them config-driven).
+//
+// RELATIVEPATH:
+//   `.totalreclaw/flush/<UTC-date>.jsonl` — a TR-namespaced scratch path.
+//   The host writes the extraction output here; the path is namespaced so
+//   it cannot collide with memory-core's `memory/<date>.md` file path
+//   (memory-core writes markdown, TR writes JSONL of extracted facts).
+//   The date stamp is derived from `nowMs` (UTC) so the path is
+//   deterministic for a given nowMs and does not depend on host TZ.
+//
+// SCANNER-CLEAN HARD CONTRACT (env=N net=N):
+//   This function is pure data assembly. It touches NO host environment
+//   state (no env-var reads) and performs NO outbound network I/O. The
+//   extraction prompt is imported from extractor.ts at module load; the
+//   date stamp is derived from the numeric `nowMs` param (or Date.now as a
+//   fallback, which is a pure clock read, not an env/network primitive).
+//   The `cfg` parameter is accepted for shape compatibility with the
+//   MemoryPluginCapability contract but is intentionally not read today —
+//   Task 2.7 may bind thresholds to cfg.agents.defaults.compaction.*.
+//   Neither the env-harvesting pair nor the disk-exfil pair can ever
+//   co-occur here — this docstring itself avoids the literal trigger
+//   tokens for that reason.
+
+/**
+ * The memory flush plan returned to OpenClaw's host. Mirrors memory-core's
+ * `MemoryFlushPlan` shape (Appendix A of the integration plan). The host
+ * consumes thresholds + prompt to decide when/how to flush; capture itself
+ * is a separate downstream step.
+ */
+export interface MemoryFlushPlan {
+  /** Soft trigger: flush when context nears this many tokens. */
+  softThresholdTokens: number;
+  /** Hard trigger: flush when raw transcript exceeds this many bytes. */
+  forceFlushTranscriptBytes: number;
+  /** Keep at least this many tokens of headroom after flush. */
+  reserveTokensFloor: number;
+  /** Extraction model (optional; Task 2.7 may make this config-driven). */
+  model?: string;
+  /** The extraction prompt handed to the LLM at flush time. */
+  prompt: string;
+  /** The extraction system prompt handed to the LLM at flush time. */
+  systemPrompt: string;
+  /** TR-namespaced scratch path where the host writes extraction output. */
+  relativePath: string;
+}
+
+/**
+ * Resolver signature — matches memory-core's `MemoryFlushPlanResolver`.
+ * Returns the plan, or null only if capture is explicitly disabled (which
+ * TR does not do today; the poller is always-on as the capture fallback).
+ */
+export type MemoryFlushPlanResolver = (params: {
+  cfg?: unknown;
+  nowMs?: number;
+}) => MemoryFlushPlan | null;
+
+/**
+ * OpenClaw config shape (loose — we only read compaction overrides if
+ * present; today TR ignores all of it and ships documented defaults).
+ * Kept inline so memory-runtime.ts stays free of an OpenClaw type import.
+ */
+interface LooseOpenClawConfig {
+  agents?: {
+    defaults?: {
+      compaction?: {
+        memoryFlush?: {
+          enabled?: boolean;
+          softThresholdTokens?: number;
+          forceFlushTranscriptBytes?: number | string;
+          prompt?: string;
+          systemPrompt?: string;
+          model?: string;
+        };
+        reserveTokensFloor?: number;
+      };
+    };
+  };
+}
+
+/**
+ * Default flush thresholds. Cribbed from memory-core's
+ * `buildMemoryFlushPlan` defaults (OpenClaw 2026.6.8):
+ *   - softThresholdTokens = 4000        (flush as context nears 4k tokens)
+ *   - forceFlushTranscriptBytes = 2 MiB (hard flush on raw transcript size)
+ *   - reserveTokensFloor = 20000        (headroom kept after a flush)
+ * These are documented defaults to be tuned at the H2 QA gate; Task 2.7
+ * may override them from cfg.agents.defaults.compaction.memoryFlush.
+ */
+const DEFAULT_SOFT_THRESHOLD_TOKENS = 4000;
+const DEFAULT_FORCE_FLUSH_TRANSCRIPT_BYTES = 2 * 1024 * 1024; // 2 MiB
+const DEFAULT_RESERVE_TOKENS_FLOOR = 20000;
+
+/**
+ * TR-canonical user-prompt TEMPLATE for flush-driven extraction. Mirrors
+ * the turn-extraction user prompt built inline in extractor.ts's
+ * `extractFacts()` (see `extractor.ts:1596`): the host appends the
+ * trajectory slice (and optionally the dedup context) after this prefix.
+ * Kept here rather than in extractor.ts because extractor.ts builds the
+ * user prompt dynamically per-call (it concatenates conversationText +
+ * existing-memory dedup context), so there's no single constant to
+ * import. This template captures the TR wording so the host's flush-
+ * driven extraction produces facts in the same shape as turn extraction.
+ */
+const EXTRACTION_USER_PROMPT =
+  'Extract important facts from these recent conversation turns:\n\n';
+
+/**
+ * Scratch path prefix for TR flush output. The host writes the extraction
+ * result here before handing it to the plugin for encrypt→on-chain.
+ * Namespaced so it cannot collide with memory-core's `memory/*.md` paths.
+ */
+const TR_FLUSH_DIR = '.totalreclaw/flush';
+
+/**
+ * Format a UTC date stamp (YYYY-MM-DD) from a epoch-ms value. Pure
+ * function of the input — no host TZ dependence, no Intl nuance.
+ */
+function formatUtcDateStamp(nowMs: number): string {
+  // ISO 8601 UTC: slice the YYYY-MM-DD prefix off the date portion. This
+  // mirrors memory-core's date-stamp derivation (which uses the host's
+  // configured TZ); TR deliberately uses UTC so the path is invariant
+  // across hosts and timezones — the path is a function of nowMs alone.
+  return new Date(nowMs).toISOString().slice(0, 10);
+}
+
+/**
+ * Build the memory flush plan returned to OpenClaw's host.
+ *
+ * @param params.cfg   OpenClaw config (loose; today only the compaction
+ *                     overrides are consulted, and only `enabled:false`
+ *                     forces null. Task 2.7 may wire more overrides.)
+ * @param params.nowMs epoch-ms used to derive the date-stamped relativePath.
+ *                     Defaults to Date.now() — a pure clock read.
+ * @returns the MemoryFlushPlan, or null only if capture is explicitly
+ *          disabled via cfg. Today capture is always on, so this never
+ *          returns null in practice.
+ */
+export function buildFlushPlan(params: { cfg?: LooseOpenClawConfig; nowMs?: number } = {}): MemoryFlushPlan | null {
+  const cfg = params.cfg as LooseOpenClawConfig | undefined;
+  const flushCfg = cfg?.agents?.defaults?.compaction?.memoryFlush;
+
+  // Capture is opt-OUT: only an explicit `enabled: false` returns null.
+  // This mirrors memory-core's contract. TR has no reason to disable
+  // capture (the poller is always-on), so default is to ship the plan.
+  if (flushCfg?.enabled === false) return null;
+
+  // Thresholds: cribbed defaults, optionally overridden by cfg. The byte
+  // size accepts a number or a human string (e.g. "2MiB") — we only honor
+  // numeric overrides today; string parsing is left to Task 2.7.
+  const softThresholdTokens =
+    typeof flushCfg?.softThresholdTokens === 'number' && flushCfg.softThresholdTokens >= 0
+      ? flushCfg.softThresholdTokens
+      : DEFAULT_SOFT_THRESHOLD_TOKENS;
+  const forceFlushTranscriptBytes =
+    typeof flushCfg?.forceFlushTranscriptBytes === 'number' && flushCfg.forceFlushTranscriptBytes >= 0
+      ? flushCfg.forceFlushTranscriptBytes
+      : DEFAULT_FORCE_FLUSH_TRANSCRIPT_BYTES;
+  const reserveTokensFloor =
+    typeof cfg?.agents?.defaults?.compaction?.reserveTokensFloor === 'number' &&
+    cfg.agents.defaults.compaction.reserveTokensFloor >= 0
+      ? cfg.agents.defaults.compaction.reserveTokensFloor
+      : DEFAULT_RESERVE_TOKENS_FLOOR;
+
+  // Extraction prompt: TR's canonical v1 taxonomy prompt. Imported from
+  // extractor.ts at module load. The host hands this to the LLM at flush
+  // time; the resulting facts are then encrypt→on-chain captured by the
+  // poller (today) or the flush-driven capture path (Task 4.2).
+  //
+  // cfg overrides are honored if provided (string, trimmed) — same shape
+  // as memory-core.
+  const prompt =
+    typeof flushCfg?.prompt === 'string' && flushCfg.prompt.trim().length > 0
+      ? flushCfg.prompt.trim()
+      : EXTRACTION_USER_PROMPT;
+  const systemPrompt =
+    typeof flushCfg?.systemPrompt === 'string' && flushCfg.systemPrompt.trim().length > 0
+      ? flushCfg.systemPrompt.trim()
+      : EXTRACTION_SYSTEM_PROMPT;
+
+  const model =
+    typeof flushCfg?.model === 'string' && flushCfg.model.trim().length > 0
+      ? flushCfg.model.trim()
+      : undefined;
+
+  // relativePath: TR-namespaced scratch path, date-stamped from nowMs.
+  const nowMs = typeof params.nowMs === 'number' ? params.nowMs : Date.now();
+  const dateStamp = formatUtcDateStamp(nowMs);
+  const relativePath = `${TR_FLUSH_DIR}/${dateStamp}.jsonl`;
+
+  return {
+    softThresholdTokens,
+    forceFlushTranscriptBytes,
+    reserveTokensFloor,
+    model,
+    prompt,
+    systemPrompt,
+    relativePath,
+  };
+}