feat(memory): LLM-assigned per-fact expected_valid_days and staleFactsToExtend (#4143)

Re-ports this feature onto the pluggable-memory backend introduced in #4122
(the original #4143 was force-pushed clean by accident and auto-closed). The
#4122 refactor moved the staleness logic into the self-contained DeerMem
backend (backends/deermem/deermem/core/) and reverted it to the pre-feature
global-threshold version, so the per-fact lifetime work is re-applied here
against the new module layout + DI (MemoryUpdater is now (config, storage,
llm)-injected; config lives on DeerMemConfig, not host MemoryConfig).

**expected_valid_days (creation)**
The LLM assigns a per-fact review window when storing each new fact. The
prompt exposes five tiers (<=14 d transient -> >365 d very stable). The value
is capped at write time by staleness_age_days x staleness_max_lifetime_multiplier
(default 20.0 -> 1800 d ~= 5 years; range 1.0-100.0) so the model cannot set
an initial lifetime so long the fact is never re-evaluated. The default 20.0
makes the "> 365 d very stable" tier achievable out of the box (3.0 silently
clamped it to 270 d).

**staleFactsToExtend (review)**
During staleness review the LLM can emit extension entries for kept facts
whose window seems miscalibrated. new_evd = min(days_since_created +
extend_by_days, staleness_max_extension_days). Extensions use an absolute
ceiling (default 3650 d ~= 10 years; range 90-36500) rather than the creation
multiplier - they are deliberate review decisions that must be able to advance
the window beyond the initial cap, but the absolute bound prevents timedelta
overflow (a model-supplied extend_by_days of 10**9 previously crashed every
later candidate-selection pass with OverflowError) and LLM misfire.

**Invariant correctness**
- Read-time cap removed from _effective_fact_staleness_age; cap is write-time
  only so extensions actually advance the review window.
- proposed_remove_ids hoisted out of the removals sub-block and used to exclude
  from extension, so a cap-surviving proposed-removal fact is never extended.
- extend_by coerced to int before the > 0 guard (a fractional 0.9 would pass
  the float check then int() to 0, silently writing a zero-delta extension).
- days_since uses total_seconds() // 86400 (not .days truncation).
- staleness-section html.escape uses quote=False to match the prompt.py
  convention; only <, >, & break element-text structure.

**Tests**
test_memory_staleness_review.py was module-level skipped by #4122 ("full
unit-test migration is a follow-up"). This PR performs that migration: DI
construction via (DeerMemConfig, _FakeStorage), _build_staleness_section back
to the (candidates, config) signature, plus new coverage for per-fact
selection, EXTEND with the absolute cap, the overflow next-cycle regression,
the proposed-removal-not-extendable case, fractional extend_by skipping, and
the creation-time cap. 67 tests, all green.

backend/AGENTS.md

@@ -580,7 +580,7 @@ The cached value is reused for both the blocking (`runs.wait`) and streaming (`_
 - `memory.mode: tool` skips `MemoryMiddleware` and registers `memory_search`, `memory_add`, `memory_update`, and `memory_delete` on the agent. The model decides when to search, add, update, or delete facts; this is opt-in/experimental and should not be described as better than middleware mode without eval evidence.
 - Both modes share `FileMemoryStorage`, per-user/per-agent isolation, prompt injection, manual CRUD primitives, and the updater backend.
 - Middleware mode queue debounces (30s default), batches updates, deduplicates per-thread, applies updates atomically (temp file + rename) with cache invalidation, and skips duplicate fact content before append.
-- Staleness pass (same LLM invocation as the regular updater, no extra API call): when `staleness_review_enabled` is `true` and at least `staleness_min_candidates` aged facts exist, `_select_stale_candidates` selects facts older than `staleness_age_days` that are not in `staleness_protected_categories` (default: `correction`), surfaces them in the prompt, and the LLM judges each as KEEP or REMOVE. `_apply_updates` enforces the guardrail unconditionally at apply time: it intersects the LLM-returned removal set with `_select_stale_candidates` output before applying the per-cycle cap (`staleness_max_removals_per_cycle`), so protected and non-aged facts can never be deleted regardless of model behavior or the feature flag setting.
+- Staleness pass (same LLM invocation as the regular updater, no extra API call): when `staleness_review_enabled` is `true` and at least `staleness_min_candidates` aged facts exist, `_select_stale_candidates` selects facts older than their individual review window (`expected_valid_days`, or the global `staleness_age_days` fallback) that are not in `staleness_protected_categories` (default: `correction`), surfaces them in the prompt with a `valid:Nd` annotation, and the LLM judges each as KEEP, REMOVE, or EXTEND. REMOVE entries go in `staleFactsToRemove`; EXTEND entries go in `staleFactsToExtend` with an `extend_by_days` value, which sets the fact's `expected_valid_days` to `min(days_since_created + extend_by_days, staleness_max_extension_days)`. The LLM assigns `expected_valid_days` when creating a fact; it is clamped at write time to `staleness_age_days × staleness_max_lifetime_multiplier` (creation cap). `_apply_updates` enforces the guardrail unconditionally at apply time: it intersects both the removal and extension sets with `_select_stale_candidates` output before applying the per-cycle cap (`staleness_max_removals_per_cycle`), so protected and non-aged facts can never be targeted regardless of model behavior or the feature flag setting. Facts the LLM proposed for removal are excluded from extension even if the per-cycle cap prevented their actual deletion that cycle. Extensions use an absolute ceiling (`staleness_max_extension_days`) rather than the creation multiplier so a deliberate review decision can advance the window beyond the initial cap while preventing `timedelta` overflow from a malformed `extend_by_days`.
 - Consolidation pass (same LLM invocation as the regular updater, no extra API call): when `consolidation_enabled` is `true` and at least one category holds `consolidation_min_facts` or more facts, `_select_consolidation_candidates` identifies fragmented categories and surfaces at most `consolidation_max_groups_per_cycle` of them (largest first) in the prompt. The LLM decides which groups to merge and proposes a synthesised fact per group. `_apply_updates` enforces guardrails: source IDs must exist and must not overlap across groups, group size is capped at `consolidation_max_sources`, the merged fact's confidence cannot exceed the source maximum, and facts below `fact_confidence_threshold` are not written.
 - Next interaction injects selected facts + context into `<memory>` tags in the system prompt when `injection_enabled` is true.
 
@@ -607,6 +607,8 @@ Focused regression coverage for the updater lives in `backend/tests/test_memory_
 - `staleness_min_candidates` - Minimum aged candidates required to trigger a review cycle (default: 3; range: 1–50)
 - `staleness_max_removals_per_cycle` - Maximum facts removed in a single cycle; lowest-confidence entries are kept when the LLM requests more (default: 10; range: 1–50)
 - `staleness_protected_categories` - Fact categories that are never pruned by staleness review (default: `["correction"]`)
+- `staleness_max_lifetime_multiplier` - Creation-time cap multiplier for a fact's LLM-assigned `expected_valid_days`: stored value is clamped to `staleness_age_days × multiplier` so the model cannot defer first review indefinitely (default: 20.0; range: 1.0–100.0). Default 20.0 (90 × 20 = 1800 d ≈ 5 years) is generous enough to support the very-stable prompt tier without needing multiple review cycles to escape the cap.
+- `staleness_max_extension_days` - Absolute upper bound (in days) on `expected_valid_days` after a lifetime extension (`staleFactsToExtend`). Applied at write time as `min(days_since + extend_by, staleness_max_extension_days)`. Uses an absolute ceiling rather than the multiplier because extensions are deliberate review decisions; prevents `timedelta` overflow and LLM misfire from permanently deferring a fact (default: 3650 = 10 years; range: 90–36500).
 - `consolidation_enabled` - Enable memory consolidation (default: `true`; no extra API call — runs in the same LLM invocation as the normal memory update)
 - `consolidation_min_facts` - Minimum facts in a category to trigger consolidation review (default: 8; range: 3–30)
 - `consolidation_max_groups_per_cycle` - Maximum categories the LLM can merge in one cycle (default: 3; range: 1–10; also controls the LLM's prompt instruction)
@@ -734,7 +736,7 @@ Config is env-driven like the others — `MonocleTracingConfig`, built in `get_t
 - `title` - Auto-title generation (enabled, max_words, max_chars, model_name; null model_name uses fast local fallback, explicit model_name uses the prompt_template LLM path)
 - `summarization` - Context summarization (enabled, trigger conditions, keep policy)
 - `subagents.enabled` - Master switch for subagent delegation
-- `memory` - Memory system (enabled, storage_path, debounce_seconds, shutdown_flush_timeout_seconds, model_name, max_facts, fact_confidence_threshold, injection_enabled, max_injection_tokens, staleness_review_enabled, staleness_age_days, staleness_min_candidates, staleness_max_removals_per_cycle, staleness_protected_categories)
+- `memory` - Memory system (enabled, storage_path, debounce_seconds, shutdown_flush_timeout_seconds, model_name, max_facts, fact_confidence_threshold, injection_enabled, max_injection_tokens, staleness_review_enabled, staleness_age_days, staleness_min_candidates, staleness_max_removals_per_cycle, staleness_protected_categories, staleness_max_lifetime_multiplier, staleness_max_extension_days)
 
 **`extensions_config.json`**:
 - `mcpServers` - Map of server name → config (enabled, type, command, args, env, url, headers, oauth, description, `routing`, `tools`, `tool_call_timeout`). `routing.mode="prefer"` emits `<mcp_routing_hints>` prompt guidance; if `tool_search` defers the hinted tool, `McpRoutingMiddleware` can also auto-promote matching deferred schemas before the model call. It does not hard-disable other tools.

backend/packages/harness/deerflow/agents/memory/backends/deermem/deermem/config.py

@@ -115,6 +115,39 @@ class DeerMemConfig(BaseModel):
         default_factory=lambda: ["correction"],
         description="Fact categories exempt from staleness review.",
     )
+    staleness_max_lifetime_multiplier: float = Field(
+        default=20.0,
+        ge=1.0,
+        le=100.0,
+        description=(
+            "Creation-time cap multiplier for a fact's LLM-assigned "
+            "expected_valid_days. When a new fact is stored, its "
+            "expected_valid_days is clamped to "
+            "staleness_age_days * staleness_max_lifetime_multiplier so the "
+            "model cannot set an initial lifetime so long that the fact is "
+            "never re-evaluated. Default 20.0 (90 x 20 = 1800 d ~= 5 years) "
+            "is generous enough to support the 'very stable' prompt tier "
+            "(core skills, native language) without needing multiple review "
+            "cycles to escape the cap. Lifetime extensions (staleFactsToExtend) "
+            "are subject to staleness_max_extension_days instead."
+        ),
+    )
+    staleness_max_extension_days: int = Field(
+        default=3650,
+        ge=90,
+        le=36500,
+        description=(
+            "Absolute upper bound (in days) on expected_valid_days after a "
+            "lifetime extension (staleFactsToExtend). Applied at write time "
+            "during staleness review: new_evd = min(days_since + extend_by, "
+            "staleness_max_extension_days). Separate from the creation-time "
+            "multiplier cap because extensions are deliberate recalibration "
+            "decisions and are not subject to the staleness_age_days scale. "
+            "The ceiling prevents a single LLM misfire from permanently "
+            "deferring a fact or causing timedelta overflow on the next "
+            "candidate-selection pass. Default 3650 (10 years)."
+        ),
+    )
     # ── Memory consolidation ────────────────────────────────────────────
     consolidation_enabled: bool = Field(
         default=False,

backend/packages/harness/deerflow/agents/memory/backends/deermem/deermem/core/prompt.py

@@ -82,6 +82,16 @@ Memory Section Guidelines:
   * behavior: Working patterns, communication habits, problem-solving approaches
   * goal: Stated objectives, learning targets, project ambitions
   * correction: Explicit agent mistakes or user corrections, including the correct approach
+- Fact lifetime (``expected_valid_days``, optional integer):
+  How many days before this fact should be reviewed for possible removal.
+  The system schedules review automatically; omit when uncertain.
+  * <= 14: highly transient - active bugs, immediate tasks, today's focus
+  * 15-60: short-term - current experiments, in-progress side projects, near-term goals
+  * 60-180: medium-term - current role, active tech stack, ongoing preferences
+  * 180-365: stable - professional background, established working patterns
+  * > 365: very stable - core skills, native language, personality traits
+  Assign the value that semantically fits; values above the server-configured
+  ceiling are silently reduced on storage.
 - Confidence levels:
   * 0.9-1.0: Explicitly stated facts ("I work on X", "My role is Y")
   * 0.7-0.8: Strongly implied from actions/discussions
@@ -114,10 +124,11 @@ Output Format (JSON):
     "longTermBackground": {{ "summary": "...", "shouldUpdate": true/false }}
   }},
   "newFacts": [
-    {{ "content": "...", "category": "preference|knowledge|context|behavior|goal|correction", "confidence": 0.0-1.0 }}
+    {{ "content": "...", "category": "preference|knowledge|context|behavior|goal|correction", "confidence": 0.0-1.0, "expected_valid_days": 90 }}
   ],
   "factsToRemove": ["fact_id_1", "fact_id_2"],
   "staleFactsToRemove": [{{ "id": "fact_id", "reason": "brief explanation" }}],
+  "staleFactsToExtend": [{{ "id": "fact_id", "extend_by_days": 365, "reason": "brief explanation" }}],
   "factsToConsolidate": [
     {{
       "sourceIds": ["fact_id_1", "fact_id_2"],
@@ -155,27 +166,37 @@ Return ONLY valid JSON, no explanation or markdown."""
 # rather than relying on passive contradiction from the current conversation.
 STALENESS_REVIEW_PROMPT = """## Staleness Review
 
-The following facts were created more than {age_days} days ago and may no longer
-accurately reflect the user's current situation. Review each one against the full
-conversation context and your understanding of the user.
+The following facts have reached their individual review window and may no longer
+accurately reflect the user's current situation. Each entry shows a ``valid:Nd``
+annotation - the number of days this fact was expected to remain valid before
+re-evaluation. Use it to calibrate conservatism: a ``valid:30d`` fact was
+considered volatile at creation; a ``valid:365d`` fact was considered stable.
 
 <stale_facts>
 {stale_facts}
 </stale_facts>
 
-For each fact, decide KEEP or REMOVE:
-- KEEP: Still likely valid — even if not mentioned in this conversation.
+For each fact, decide KEEP, REMOVE, or EXTEND:
+- KEEP: Still likely valid - even if not mentioned in this conversation.
   Stable attributes (native language, core expertise, personality traits) often
   remain true indefinitely.
 - REMOVE: Outdated, contradicted by recent context, or no longer relevant.
   Examples: tech-stack migrations, job changes, relocated offices, abandoned projects.
+- EXTEND: Keep but recalibrate the review window (see below).
 
 Add REMOVE decisions to "staleFactsToRemove" in your output JSON.
 Each entry must be {{"id": "fact_id", "reason": "brief explanation"}}.
 The reason should cite what signal in the conversation (or absence thereof)
 supports the removal.
 
-Be conservative — when in doubt, KEEP. Removing a valid fact is worse than
+Optionally, for facts you KEEP and wish to recalibrate, add them to
+"staleFactsToExtend" with the number of days from now before the next review:
+{{"id": "fact_id", "extend_by_days": 365, "reason": "brief explanation"}}
+Use this when the current window seems miscalibrated - e.g. a core skill marked
+``valid:30d`` that is clearly stable, or a goal nearing completion whose window
+should shrink. Omit facts whose current window already seems appropriate.
+
+Be conservative - when in doubt, KEEP. Removing a valid fact is worse than
 keeping a slightly stale one, because the next review cycle will re-evaluate it."""
 
 

backend/packages/harness/deerflow/agents/memory/backends/deermem/deermem/core/updater.py

@@ -171,6 +171,15 @@ def _normalize_memory_update_fact(fact: Any) -> dict[str, Any] | None:
         if normalized_source_error:
             normalized_fact["sourceError"] = normalized_source_error
 
+    # Fact lifetime (expected_valid_days): optional LLM-assigned review window.
+    # Accept int/float (reject bool which subclasses int), coerce to int, keep
+    # only positive values; the creation-time cap is applied in _apply_updates.
+    raw_evd = fact.get("expected_valid_days")
+    if isinstance(raw_evd, (int, float)) and not isinstance(raw_evd, bool):
+        evd = int(raw_evd)
+        if evd > 0:
+            normalized_fact["expected_valid_days"] = evd
+
     return normalized_fact
 
 
@@ -216,6 +225,32 @@ def _normalize_memory_update_data(update_data: dict[str, Any]) -> dict[str, Any]
                 }
             )
 
+    # ── Normalize staleness review lifetime extensions ──
+    stale_extensions_raw = update_data.get("staleFactsToExtend")
+    normalized_stale_extensions: list[dict[str, Any]] = []
+    if isinstance(stale_extensions_raw, list):
+        for entry in stale_extensions_raw:
+            if not isinstance(entry, dict):
+                continue
+            fact_id = entry.get("id")
+            if not isinstance(fact_id, str) or not fact_id:
+                continue
+            # extend_by_days: accept int/float (reject bool), coerce to int, keep > 0.
+            # A fractional value in (0, 1) coerces to 0 and is dropped here so the
+            # apply path never silently writes a zero-delta extension.
+            raw_extend = entry.get("extend_by_days")
+            if isinstance(raw_extend, (int, float)) and not isinstance(raw_extend, bool):
+                extend_by = int(raw_extend)
+                if extend_by > 0:
+                    reason = entry.get("reason", "")
+                    normalized_stale_extensions.append(
+                        {
+                            "id": fact_id,
+                            "extend_by_days": extend_by,
+                            "reason": reason if isinstance(reason, str) else "",
+                        }
+                    )
+
     # ── Normalize consolidation decisions ──
     consolidation_raw = update_data.get("factsToConsolidate")
     normalized_consolidation: list[dict[str, Any]] = []
@@ -265,6 +300,7 @@ def _normalize_memory_update_data(update_data: dict[str, Any]) -> dict[str, Any]
         "newFacts": normalized_new_facts,
         "factsToRemove": normalized_facts_to_remove,
         "staleFactsToRemove": normalized_stale_removals,
+        "staleFactsToExtend": normalized_stale_extensions,
         "factsToConsolidate": normalized_consolidation,
     }
 
@@ -357,16 +393,38 @@ def _parse_fact_datetime(raw: str) -> datetime | None:
         return None
 
 
+def _effective_fact_staleness_age(fact: dict[str, Any], config: Any) -> int:
+    """Return the effective staleness review age in days for *fact*.
+
+    Returns the stored ``expected_valid_days`` value directly when present and
+    valid.  The ``staleness_max_lifetime_multiplier`` cap is applied once at
+    *write time* (when a fact is first created) so the review window is bounded
+    from the start.  Re-applying it here would prevent lifetime-extension
+    operations from ever moving the review window beyond that initial cap,
+    defeating the purpose of ``staleFactsToExtend``.  Falls back to the global
+    ``staleness_age_days`` for facts that pre-date this feature or where the
+    LLM did not provide an estimate.
+    """
+    raw = fact.get("expected_valid_days")
+    if isinstance(raw, (int, float)) and not isinstance(raw, bool) and raw > 0:
+        return int(raw)
+    return config.staleness_age_days
+
+
 def _select_stale_candidates(
     current_memory: dict[str, Any],
     config: Any,
 ) -> list[dict[str, Any]]:
-    """Return facts that are older than ``staleness_age_days`` and not protected.
+    """Return facts that have exceeded their individual review window.
 
-    Protected categories (default: ``correction``) are excluded because they
-    represent explicit user feedback that should not be auto-pruned by age.
+    Each fact's effective review age is determined by
+    ``_effective_fact_staleness_age``: facts with an LLM-assigned
+    ``expected_valid_days`` use that value directly; facts without it fall back
+    to the global ``staleness_age_days``.  Protected categories (default:
+    ``correction``) are excluded because they represent explicit user feedback
+    that should not be auto-pruned by age.
     """
-    cutoff = datetime.now(UTC) - timedelta(days=config.staleness_age_days)
+    now = datetime.now(UTC)
     protected = frozenset(config.staleness_protected_categories)
     candidates: list[dict[str, Any]] = []
     for fact in current_memory.get("facts", []):
@@ -376,31 +434,41 @@ def _select_stale_candidates(
         if isinstance(category, str) and category in protected:
             continue
         created_at = _parse_fact_datetime(fact.get("createdAt", ""))
-        if created_at is not None and created_at < cutoff:
+        if created_at is None:
+            continue
+        effective_age = _effective_fact_staleness_age(fact, config)
+        if created_at < now - timedelta(days=effective_age):
             candidates.append(fact)
     return candidates
 
 
 def _build_staleness_section(
     stale_candidates: list[dict[str, Any]],
-    age_days: int,
+    config: Any,
 ) -> str:
-    """Format the staleness review prompt section from candidate facts."""
+    """Format the staleness review prompt section from candidate facts.
+
+    Each fact line includes a ``valid:Nd`` annotation - the effective review
+    window for that fact - so the LLM can calibrate its conservatism: a fact
+    reviewed after 30 days was considered volatile at creation; one reviewed
+    after 365 days was considered stable.
+    """
     if not stale_candidates:
         return ""
     lines: list[str] = []
     for fact in stale_candidates:
         fid = fact.get("id", "?")
-        cat = html.escape(str(fact.get("category", "context")).strip() or "context")
+        cat = html.escape(str(fact.get("category", "context")).strip() or "context", quote=False)
         conf = _coerce_source_confidence(fact)
         created_raw = fact.get("createdAt", "")
         created_short = created_raw[:10] if isinstance(created_raw, str) and len(created_raw) >= 10 else created_raw
-        content = html.escape(str(fact.get("content", "")))
-        lines.append(f'- [{fid} | {cat} | {conf:.2f} | {created_short}] "{content}"')
-    return STALENESS_REVIEW_PROMPT.format(
-        stale_facts="\n".join(lines),
-        age_days=age_days,
-    )
+        # quote=False: content is in element-text position (inside <stale_facts>
+        # tags, never an attribute value), so only <, >, & can break structure -
+        # leave ' and " untouched. Mirrors the convention in prompt.py #4028.
+        content = html.escape(str(fact.get("content", "")), quote=False)
+        effective_age = _effective_fact_staleness_age(fact, config)
+        lines.append(f'- [{fid} | {cat} | {conf:.2f} | {created_short} | valid:{effective_age}d] "{content}"')
+    return STALENESS_REVIEW_PROMPT.format(stale_facts="\n".join(lines))
 
 
 # ── Consolidation helpers ───────────────────────────────────────────────
@@ -670,10 +738,7 @@ class MemoryUpdater:
         if config.staleness_review_enabled:
             stale_candidates = _select_stale_candidates(current_memory, config)
             if len(stale_candidates) >= config.staleness_min_candidates:
-                staleness_section = _build_staleness_section(
-                    stale_candidates,
-                    config.staleness_age_days,
-                )
+                staleness_section = _build_staleness_section(stale_candidates, config)
 
         # ── Build consolidation section ──
         consolidation_section = ""
@@ -953,48 +1018,101 @@ class MemoryUpdater:
         if facts_to_remove:
             current_memory["facts"] = [f for f in current_memory.get("facts", []) if f.get("id") not in facts_to_remove]
 
-        # ── Staleness review removals ──
+        # ── Staleness review: removals + lifetime extensions ──
+        # Both operations share one staleness-candidate guardrail pass and one
+        # candidate_ids set. proposed_remove_ids is hoisted out of the removals
+        # sub-block so it covers ALL LLM-proposed removals, not just the ones
+        # the per-cycle cap actually deleted: a fact the LLM wanted to remove
+        # must never be silently extended even when the cap spares it.
         stale_removals = update_data.get("staleFactsToRemove", [])
-        if isinstance(stale_removals, list) and stale_removals:
-            stale_ids_to_remove = {entry["id"] for entry in stale_removals if isinstance(entry, dict) and "id" in entry}
-
-            # Deterministic guardrail: intersect with actual staleness
-            # candidates so an LLM slip that emits a protected-category or
-            # non-aged fact id is silently rejected.  Runs unconditionally
-            # so the apply-layer protection is independent of model behavior
-            # AND of the staleness_review_enabled flag.
-            # Guard against legacy / hand-edited facts that predate the id
-            # field: an aged, non-protected fact with no "id" is a valid
-            # staleness candidate but has no id to intersect against, so skip
-            # it here instead of raising KeyError (id-less facts can never be
-            # targeted by the id-based removal set anyway).
+        stale_extensions = update_data.get("staleFactsToExtend", [])
+        has_staleness_ops = (isinstance(stale_removals, list) and stale_removals) or (isinstance(stale_extensions, list) and stale_extensions)
+        if has_staleness_ops:
+            # Deterministic guardrail: intersect with actual staleness candidates
+            # so an LLM slip that emits a protected-category or non-aged fact id
+            # is silently rejected.  Runs unconditionally so the apply-layer
+            # protection is independent of model behavior AND of the
+            # staleness_review_enabled flag.  Guard against legacy / hand-edited
+            # facts that predate the id field: an aged, non-protected fact with
+            # no "id" is a valid staleness candidate but has no id to intersect
+            # against, so skip it here instead of raising KeyError.
             candidate_ids = {f["id"] for f in _select_stale_candidates(current_memory, config) if f.get("id") is not None}
-            stale_ids_to_remove &= candidate_ids
 
-            if not stale_ids_to_remove:
-                # After intersection with candidate set, nothing to remove.
-                stale_removals = []
-            else:
-                # Safety cap: limit max staleness removals per cycle.
-                # When the LLM returns more than the cap, keep only the
-                # lowest-confidence entries up to the limit so the most
-                # questionable facts are removed first.
-                max_stale = config.staleness_max_removals_per_cycle
-                if len(stale_ids_to_remove) > max_stale:
-                    stale_facts = [f for f in current_memory.get("facts", []) if f.get("id") in stale_ids_to_remove]
-                    stale_facts.sort(key=_coerce_source_confidence)
-                    stale_ids_to_remove = {f["id"] for f in stale_facts[:max_stale]}
+            # ── Removals ──
+            proposed_remove_ids: set[str] = set()
+            if isinstance(stale_removals, list) and stale_removals:
+                proposed_remove_ids = {entry["id"] for entry in stale_removals if isinstance(entry, dict) and "id" in entry}
+                stale_ids_to_remove = proposed_remove_ids & candidate_ids
 
-                current_memory["facts"] = [f for f in current_memory.get("facts", []) if f.get("id") not in stale_ids_to_remove]
+                if not stale_ids_to_remove:
+                    stale_removals = []
+                else:
+                    # Safety cap: limit max staleness removals per cycle.  When
+                    # the LLM returns more than the cap, keep only the
+                    # lowest-confidence entries up to the limit so the most
+                    # questionable facts are removed first.
+                    max_stale = config.staleness_max_removals_per_cycle
+                    if len(stale_ids_to_remove) > max_stale:
+                        stale_facts = [f for f in current_memory.get("facts", []) if f.get("id") in stale_ids_to_remove]
+                        stale_facts.sort(key=_coerce_source_confidence)
+                        stale_ids_to_remove = {f["id"] for f in stale_facts[:max_stale]}
 
-            # Log removals for observability
-            for entry in stale_removals:
-                if isinstance(entry, dict) and entry.get("id") in stale_ids_to_remove:
-                    logger.info(
-                        "Staleness review removed fact %s: %s",
-                        entry["id"],
-                        entry.get("reason", "no reason provided"),
-                    )
+                    current_memory["facts"] = [f for f in current_memory.get("facts", []) if f.get("id") not in stale_ids_to_remove]
+
+                # Log removals for observability
+                for entry in stale_removals:
+                    if isinstance(entry, dict) and entry.get("id") in stale_ids_to_remove:
+                        logger.info(
+                            "Staleness review removed fact %s: %s",
+                            entry["id"],
+                            entry.get("reason", "no reason provided"),
+                        )
+
+            # ── Lifetime extensions ──
+            # Recalibrate expected_valid_days for facts the LLM chose to keep.
+            # Eligible facts are stale candidates that the LLM did NOT propose
+            # for removal - including those that survived only because the
+            # per-cycle cap prevented their deletion.  The new window is
+            # min(days_since + extend_by_days, staleness_max_extension_days).
+            # Extensions use an absolute ceiling rather than the creation-time
+            # multiplier cap: they are deliberate review decisions and must be
+            # able to advance the window beyond the original creation cap, but
+            # an absolute bound prevents timedelta overflow and LLM misfire.
+            if isinstance(stale_extensions, list) and stale_extensions:
+                # Exclude all LLM-proposed removals, not just the trimmed set,
+                # so a cap-surviving proposed-removal fact is never extended.
+                extendable_ids = candidate_ids - proposed_remove_ids
+                ext_by_id = {e["id"]: e for e in stale_extensions if isinstance(e, dict) and isinstance(e.get("id"), str) and e["id"] in extendable_ids}
+                if ext_by_id:
+                    now_utc = datetime.now(UTC)
+                    max_ext = config.staleness_max_extension_days
+                    updated_facts: list[dict[str, Any]] = []
+                    for fact in current_memory.get("facts", []):
+                        fid = fact.get("id")
+                        ext = ext_by_id.get(fid) if fid else None
+                        if ext is not None:
+                            extend_by = ext.get("extend_by_days")
+                            if isinstance(extend_by, (int, float)) and not isinstance(extend_by, bool):
+                                extend_by_int = int(extend_by)  # coerce before guard
+                                if extend_by_int > 0:
+                                    created = _parse_fact_datetime(fact.get("createdAt", ""))
+                                    if created is None:
+                                        # Unreachable: _select_stale_candidates already
+                                        # excludes facts with unparseable createdAt.
+                                        updated_facts.append(fact)
+                                        continue
+                                    days_since = int((now_utc - created).total_seconds() // 86400)
+                                    new_evd = min(days_since + extend_by_int, max_ext)
+                                    fact = {**fact, "expected_valid_days": new_evd}
+                                    logger.info(
+                                        "Staleness review extended fact %s by %d days (new expected_valid_days: %d): %s",
+                                        fid,
+                                        extend_by_int,
+                                        new_evd,
+                                        ext.get("reason", "no reason provided"),
+                                    )
+                        updated_facts.append(fact)
+                    current_memory["facts"] = updated_facts
 
         # Add new facts
         existing_fact_keys = {fact_key for fact_key in (_fact_content_key(fact.get("content")) for fact in current_memory.get("facts", [])) if fact_key is not None}
@@ -1028,6 +1146,15 @@ class MemoryUpdater:
                     normalized_source_error = source_error.strip()
                     if normalized_source_error:
                         fact_entry["sourceError"] = normalized_source_error
+                evd = fact.get("expected_valid_days")
+                if isinstance(evd, int) and not isinstance(evd, bool) and evd > 0:
+                    # Apply the creation-time cap so the LLM cannot assign an
+                    # unbounded lifetime that defers staleness review indefinitely.
+                    # Extensions (staleFactsToExtend) bypass this cap via their own
+                    # staleness_max_extension_days ceiling because they represent a
+                    # deliberate review decision, not an unchecked initial assignment.
+                    creation_cap = int(config.staleness_age_days * config.staleness_max_lifetime_multiplier)
+                    fact_entry["expected_valid_days"] = min(evd, creation_cap)
                 current_memory["facts"].append(fact_entry)
                 if fact_key is not None:
                     existing_fact_keys.add(fact_key)

backend/packages/harness/deerflow/config/memory_config.py

@@ -41,6 +41,8 @@ _LEGACY_DEERMEM_FIELDS = frozenset(
         "staleness_min_candidates",
         "staleness_max_removals_per_cycle",
         "staleness_protected_categories",
+        "staleness_max_lifetime_multiplier",
+        "staleness_max_extension_days",
         "consolidation_enabled",
         "consolidation_min_facts",
         "consolidation_max_groups_per_cycle",

config.example.yaml

@@ -1474,6 +1474,15 @@ summarization:
 #   staleness_min_candidates       - Minimum stale facts to trigger review (default: 3)
 #   staleness_max_removals_per_cycle - Safety cap on removals per cycle (default: 10)
 #   staleness_protected_categories - Categories exempt from staleness review (default: ["correction"])
+#   staleness_max_lifetime_multiplier - Creation-time cap multiplier for a fact LLM-assigned
+#                                       expected_valid_days; new facts clamped to
+#                                       staleness_age_days x multiplier. Default 20.0
+#                                       (90 x 20 = 1800d ~ 5 years) supports the very-stable
+#                                       prompt tier. (default: 20.0, range: 1.0-100.0)
+#   staleness_max_extension_days   - Absolute ceiling (in days) on expected_valid_days after a
+#                                    lifetime extension (staleFactsToExtend). Prevents timedelta
+#                                    overflow and LLM misfire from permanently deferring a fact.
+#                                    (default: 3650, range: 90-36500)
 memory:
   enabled: true
   injection_enabled: true # Whether to inject memory into system prompt
@@ -1527,14 +1536,19 @@ memory:
     guaranteed_token_budget: 500
     # Staleness review: periodically prune aged facts that may no longer reflect
     # the user current situation. When triggered, the LLM reviews facts older
-    # than staleness_age_days during the normal memory-update call (same LLM
-    # invocation - no extra API call) and decides KEEP or REMOVE for each.
+    # than their individual review window (expected_valid_days, or
+    # staleness_age_days as fallback) during the normal memory-update call (same
+    # LLM invocation - no extra API call) and decides KEEP, REMOVE, or EXTEND
+    # for each. The LLM assigns expected_valid_days when creating a fact; EXTEND
+    # (staleFactsToExtend) recalibrates that window at review time.
     staleness_review_enabled: true
     staleness_age_days: 90
     staleness_min_candidates: 3
     staleness_max_removals_per_cycle: 10
     staleness_protected_categories:
       - correction
+    staleness_max_lifetime_multiplier: 20.0   # creation cap = staleness_age_days x multiplier (90 x 20 = 1800d)
+    staleness_max_extension_days: 3650        # absolute ceiling on extended expected_valid_days (~10 years)
     # Memory consolidation (opt-in, lossy: source facts are replaced by a
     # synthesized one; only consolidatedFrom IDs are kept). Runs in the same
     # memory-update LLM call as extraction/staleness - no extra API cost.