Tianye Song 8da7cbf028
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.
2026-07-16 09:34:21 +08:00

200 lines
8.8 KiB
Python

"""Configuration for the memory mechanism (host-shared fields only).
DeerMem-private fields live in ``backends/deermem/config.py`` (``DeerMemConfig``),
reached via ``backend_config`` (a dict the factory passes to the backend's
``__init__``). This module holds ONLY the host-shared fields every backend /
call site / factory reads: ``enabled`` / ``injection_enabled`` /
``shutdown_flush_timeout_seconds`` / ``manager_class`` / ``backend_config``.
Keeping the shared schema slim is what
makes backends swappable and portable (DeerMem's knobs do not leak onto the
shared contract).
"""
import logging
from typing import Any, Literal
from pydantic import BaseModel, Field
logger = logging.getLogger(__name__)
# Host-shared MemoryConfig fields (read by every backend / call site / factory).
_SHARED_FIELDS = frozenset({"enabled", "mode", "injection_enabled", "shutdown_flush_timeout_seconds", "manager_class", "backend_config"})
# DeerMem-private fields that used to live at the top level of `memory:` in
# config.yaml (pre-abstraction). On load they are auto-migrated into
# `backend_config` so an upgrade does NOT silently revert customized settings
# to defaults. `model_name` maps to `backend_config.model.model` (the new nested
# model sub-config); the rest are 1:1.
_LEGACY_DEERMEM_FIELDS = frozenset(
{
"storage_path",
"storage_class",
"debounce_seconds",
"max_facts",
"fact_confidence_threshold",
"max_injection_tokens",
"token_counting",
"guaranteed_categories",
"guaranteed_token_budget",
"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",
"consolidation_enabled",
"consolidation_min_facts",
"consolidation_max_groups_per_cycle",
"consolidation_max_sources",
"model_name",
}
)
class MemoryConfig(BaseModel):
"""Host-shared memory configuration (backend-agnostic)."""
enabled: bool = Field(
default=True,
description="Whether to enable the memory mechanism (call-site gate).",
)
mode: Literal["middleware", "tool"] = Field(
default="middleware",
description=(
"Memory operation mode. 'middleware': passive LLM summarization after each turn (current behavior). 'tool': model calls memory tools (memory_search, memory_add, etc.) directly. Mutually exclusive — only one mode runs at a time."
),
)
injection_enabled: bool = Field(
default=True,
description="Whether to inject memory into the system prompt (call-site gate).",
)
shutdown_flush_timeout_seconds: float = Field(
default=30.0,
ge=1.0,
le=300.0,
description=(
"Hard time budget (seconds) for draining the memory backend's "
"pending-update buffer during Gateway graceful shutdown. The drain "
"makes one LLM call per pending item, so large IM batches may need "
"a higher value. Must fit inside the pod's K8s "
"terminationGracePeriodSeconds (together with channel/scheduler "
"stop) or K8s SIGKILLs the drain mid-flight. The drain runs on a "
"daemon thread, so on timeout the process proceeds to exit and any "
"unfinished tail is dropped (same failure direction as no flush, "
"scoped to the tail). Host-shared (not backend-private): the host "
"owns the lifespan budget and the K8s grace relationship."
),
)
manager_class: str = Field(
default="deermem",
description=(
"Memory backend selector. Either a registered backend name "
"(matching a `backends/<name>/` folder that exposes `MANAGER_CLASS`, "
"e.g. `deermem` / `noop`) or a dotted import path to a "
"`MemoryManager` subclass. The factory resolves this at "
"`get_memory_manager()` time and raises `ValueError` on failure "
"(fail-fast: memory is persistent state, so an unresolved "
"manager_class is not silently substituted with a different "
"storage backend)."
),
)
backend_config: dict[str, Any] = Field(
default_factory=dict,
description=(
"Backend-private config (a dict), passed verbatim to the backend's "
"`__init__(backend_config=...)` by the factory. Each backend "
"self-interprets it (DeerMem parses it into `DeerMemConfig`). Values "
"live in the host config file (`config.yaml` `memory.backend_config`); "
"they do not belong on the shared `MemoryConfig` schema."
),
)
def should_use_memory_tools(config: MemoryConfig) -> bool:
"""Return True when memory should use model-directed tools."""
return config.enabled and config.mode == "tool"
# Global configuration instance
_memory_config: MemoryConfig = MemoryConfig()
def get_memory_config() -> MemoryConfig:
"""Get the current memory configuration."""
return _memory_config
def set_memory_config(config: MemoryConfig) -> None:
"""Set the memory configuration."""
global _memory_config
_memory_config = config
def load_memory_config_from_dict(config_dict: dict) -> None:
"""Load memory configuration from a dictionary.
Host-shared fields (``enabled`` / ``mode`` / ``injection_enabled`` /
``manager_class`` / ``backend_config``) are read directly. DeerMem-private
fields that used to live at the top level of ``memory:`` in config.yaml
(pre-abstraction: ``storage_path``, ``max_facts``, ``debounce_seconds``,
``model_name``, ``token_counting``, ``staleness_*``, ``consolidation_*``,
...) are **auto-migrated into ``backend_config``** with a warning, so an
upgrade from a pre-abstraction config does NOT silently revert customized
settings to defaults. Unknown top-level keys (likely typos) are warned and
ignored.
"""
global _memory_config
config_dict = dict(config_dict or {})
backend_config = dict(config_dict.get("backend_config") or {})
migrated: list[str] = []
for key in list(config_dict.keys()):
if key in _SHARED_FIELDS:
continue
if key in _LEGACY_DEERMEM_FIELDS:
value = config_dict.pop(key)
if value is None or value == "":
continue # default / empty value, no migration needed
if key == "model_name":
# old top-level model_name -> backend_config.model.model
model_cfg = dict(backend_config.get("model") or {})
if "model" not in model_cfg:
model_cfg["model"] = value
backend_config["model"] = model_cfg
migrated.append(f"{key} -> backend_config.model.model")
elif key == "storage_path" and str(value).endswith(".json"):
# Pre-abstraction storage_path was a FILE path (absolute = shared
# file opting out of per-user; a relative value like the old default
# "memory.json" was ignored for per-user). DeerMem now treats it as a
# root DIRECTORY. Carrying a file-style value verbatim would be
# resolved as a dir and either orphan per-user memory or hit
# NotADirectoryError on save. Drop it so the factory's zero-config
# runtime_home kicks in (per-user location unchanged:
# {base_dir}/users/{uid}/memory.json) and warn the operator.
logger.warning(
"Legacy memory.storage_path=%r looks like a file path; DeerMem now "
"treats storage_path as a root DIRECTORY (per-user memory under "
"{storage_path}/users/{uid}/memory.json). Dropped -- memory now "
"lands under the default root (runtime_home). Set "
"memory.backend_config.storage_path to a directory if you want a "
"custom location.",
value,
)
elif key not in backend_config:
# don't override an explicit backend_config value
backend_config[key] = value
migrated.append(f"{key} -> backend_config.{key}")
else:
logger.warning(
"Unknown memory config key %r at top level (not a shared field %s nor a known legacy DeerMem field); ignored.",
key,
sorted(_SHARED_FIELDS),
)
if migrated:
logger.warning(
"Migrated legacy top-level memory fields into backend_config; move them under memory.backend_config in config.yaml to silence this: %s",
", ".join(migrated),
)
config_dict["backend_config"] = backend_config
_memory_config = MemoryConfig(**config_dict)