feat: add Keenable as a configurable internet search backend

[ilya-bogin-keenable]

Summary

Adds Keenable as a new internet search backend alongside the existing Bocha / Tavily / Google / Bing / Xinyu retrievers, following the Tavily backend pattern (#1357). Additive and opt-in via INTERNET_SEARCH_BACKEND=keenable; existing backends are untouched.

Keenable is a web search API built for AI agents. Unlike the key-required backends it is keyless by default: with no key it calls the public endpoint (rate-limited), and an optional KEENABLE_API_KEY only lifts the cap.

Files changed

• src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (new): InternetKeenableRetriever. No SDK dependency, a thin requests call. Keyless requests hit /v1/search/public; a configured key switches to /v1/search with an X-API-Key header. Attribution via X-Keenable-Title. Results map into TextualMemoryItem the same way the Tavily retriever does.
• src/memos/configs/internet_retriever.py: KeenableSearchConfig (API key optional) + registration in InternetRetrieverConfigFactory.
• src/memos/memories/textual/tree_text_memory/retrieve/internet_retriever_factory.py: register keenable + constructor branch.
• src/memos/api/config.py: INTERNET_SEARCH_BACKEND=keenable branch (KEENABLE_API_KEY optional).

Testing

• python -m py_compile on all changed files: passes.
• ruff check on all changed files: passes.

[ilya-bogin-keenable]

Hey @bittergreen could you please take a look? Thanks!

[Memtensor-AI]

🤖 Open Code Review

Target: PR #2072
Task: e42b97dfdcb50f4d
Base: main
Head: feat/keenable-web-search

🔍 OpenCodeReview found 8 issue(s) in this PR.

---

1. src/memos/configs/internet_retriever.py (L90-L92)

The api_key field in KeenableSearchConfig overrides the base class BaseInternetRetrieverConfig.api_key: str = Field(...) (required, non-nullable) with str | None and a None default. This is a Liskov Substitution Principle violation: any code that receives a BaseInternetRetrieverConfig and accesses .api_key expecting a str will silently get None for Keenable instances, which can cause AttributeError/TypeError at runtime in generic consumers.

Consider either:

1. Widening the base class declaration to api_key: str | None = Field(default=None, ...) (consistent with how search_engine_id is already declared), or
2. Keeping the override but adding a Pydantic model_validator / @field_validator in KeenableSearchConfig to document and enforce the None-is-valid contract explicitly.

💡 Suggested Change

Before:

    api_key: str | None = Field(
        default=None, description="Keenable API key (optional; keyless by default)"
    )

After:

# Option 1: widen the base class (preferred, keeps LSP intact)
# In BaseInternetRetrieverConfig:
#   api_key: str | None = Field(default=None, description="API key for the search service")

# Option 2: keep the override, add an explicit validator in KeenableSearchConfig
    api_key: str | None = Field(
        default=None, description="Keenable API key (optional; keyless by default)"
    )

    @field_validator("api_key", mode="before")
    @classmethod
    def _coerce_api_key(cls, v: str | None) -> str | None:
        return (v or "").strip() or None

---

2. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L195-L199)

The broad except Exception silently swallows all failures — connection errors, timeouts, HTTP 4xx/5xx, and JSON decode errors — and returns an empty list indistinguishable from a legitimate zero-result response. The caller in internet_retriever_factory.py has no way to distinguish a real error from an empty result set, so failures are invisible at the application level.

Split the try block so each failure point is guarded separately, and either re-raise or use a typed sentinel so the caller can react:

try:
    resp = requests.post(...)
    resp.raise_for_status()
except requests.exceptions.Timeout:
    logger.error("Keenable search timed out for query: %s", query)
    return []
except requests.exceptions.RequestException as e:
    logger.error("Keenable search request failed: %s", e)
    return []

try:
    raw_results = resp.json().get("results", [])[:limit]
except ValueError as e:
    logger.error("Keenable search returned invalid JSON: %s", e)
    return []

This makes the origin of each failure obvious and avoids accidentally catching unrelated exceptions like MemoryError or KeyboardInterrupt.

---

3. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L60)

TextRank is instantiated once at __init__ time and then called concurrently from multiple threads via ContextThreadPoolExecutor in _convert_to_mem_items. jieba's TextRank maintains internal mutable state (segment caches, graph structures), so concurrent textrank() calls on the same instance can produce incorrect results or raise exceptions under load.

Either create a new TextRank() instance per call inside _process_result, or protect the shared instance with a threading.Lock:

from threading import Lock
# in __init__:
self._textrank_lock = Lock()

# in _process_result:
if lang == "zh":
    with self._textrank_lock:
        tags = self.zh_fast_keywords_extractor.textrank(summary, topK=3)[:3]

---

4. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L297)

embed([content])[0] assumes the returned list is non-empty. If embed() returns an empty list for any reason (internal error, tokenization producing zero tokens, etc.), this raises an IndexError inside the thread pool, which is then silently swallowed by the except Exception as e in _convert_to_mem_items.

Guard the call:

embeddings = self.embedder.embed([content]) if content else []
embedding = embeddings[0] if embeddings else []

---

5. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L254-L255)

The except Exception around the ISO datetime parse is broader than necessary. Only ValueError (invalid format string) is expected here; catching all Exception types masks programming errors like AttributeError if publish_time is not a string. Narrow it:

except (ValueError, TypeError):
    publish_time = datetime.now().strftime("%Y-%m-%d")

---

6. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L237)

Deduplication keyed on item.memory (the full memory_text string) is fragile. Two results from the same URL with minor whitespace or date differences will both survive, while two different articles that happen to produce the same formatted string will be incorrectly merged. URL is a more stable and semantically correct deduplication key:

def _url(item):
    sources = item.metadata.sources
    return sources[0].url if sources else item.memory

unique_memory_items = {_url(item): item for item in memory_items}

---

7. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L240-L242)

parsed_goal is annotated as str but is used as an object with a .tags attribute in _extract_tags (hasattr(parsed_goal, 'tags')). The type annotation is incorrect and misleading. Change it to Any or the actual goal type used in the rest of the codebase:

def _process_result(
    self, result: dict, query: str, parsed_goal: Any, info: dict[str, Any], mode="fast"
) -> list[TextualMemoryItem]:

---

8. src/memos/memories/textual/tree_text_memory/retrieve/keenablesearch.py (L196-L198)

import traceback inside the except block is unconventional; it should be a top-level module import. Additionally, logger.error(f"...") eagerly formats the string even when the error level is filtered. Prefer the lazy %s style:

import traceback  # at the top of the file

# in the except block:
logger.error("Keenable search error: %s", traceback.format_exc())

Generated by cloud-assistant via Open Code Review.

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 4s [advisory, non-gating] AI-generated tests on branch test/auto-gen-33586be3c2b3f89c-20260801031509: 107/107 passed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[ilya-bogin-keenable]

Аixed 4 of the 5: jieba is now gated behind require_python_package like the bocha retriever (an English-only install used to fail at construction), empty content is no longer embedded, and GDP/AI are lowercased so they can actually match.

Will not fix for #2. Forwarding the local mode to the API would break the default path: the search endpoint returns 400 for mode: "fast", and fast is the default the caller passes. The API's mode and this method's mode are different things, so I added a comment saying so instead.

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 4s [advisory, non-gating] AI-generated tests on branch test/auto-gen-b3c50e54d556e024-20260804132254: 86/87 passed, 1 failed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 4s [advisory, non-gating] AI-generated tests on branch test/auto-gen-d929796b8f2bd83f-20260819141815: 136/136 passed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[Memtensor-AI]

⚠️ Automated Test Results: ENV ISSUE

The test environment encountered an issue that requires manual attention.

Details: Executor error: Command failed: git clone --depth 1 --branch feat/keenable-web-search git@github.com:keenableai/MemOS.git /data/test-workspaces/0bb2eb3c8566c1ed/repo
Cloning into '/data/test-workspaces/0bb2eb3c8566c1ed/repo'...
Branch: feat/keenable-web-search

[Memtensor-AI]

⚠️ Automated Test Results: ENV ISSUE

The test environment encountered an issue that requires manual attention.

Details: Executor error: Command failed: git clone --depth 1 --branch feat/keenable-web-search git@github.com:keenableai/MemOS.git /data/test-workspaces/1a3c8d0e9d69e77a/repo
Cloning into '/data/test-workspaces/1a3c8d0e9d69e77a/repo'...
kex_exchange_identification: Connection closed by remote host
Connection closed by UNKNOWN port 65535
fatal: Could not read from remote repository.

Please make sure you have the correct access rights
and the repository exists.
Branch: feat/keenable-web-search

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 5s [advisory, non-gating] AI-generated tests on branch test/auto-gen-6b2de16b7c284ae6-20260828124956: 123/125 passed, 2 failed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 4s [advisory, non-gating] AI-generated tests on branch test/auto-gen-ffe1c10a605ed5dd-20260829140808: 106/107 passed, 1 failed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 4s [advisory, non-gating] AI-generated tests on branch test/auto-gen-aa46054cf1f71848-20260901114424: 90/92 passed, 2 failed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (5/5 executed). memos_python_core/changed-python-source: 5/5. Duration: 4s [advisory, non-gating] AI-generated tests on branch test/auto-gen-e42b97dfdcb50f4d-20260902191710: 109/111 passed, 2 failed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/keenable-web-search

[ilya-bogin-keenable]

@WeiminLee @syzsunshine219 this has been approved twice and MERGEABLE since 2026-08-01, and the approval has now survived fourteen base merges; the diff is unchanged at 4 files, +346/-1.

I would rather stop refreshing it than keep it warm indefinitely, so any of three answers ends this for me: someone merges it, someone requests changes, or you tell me MemOS does not want a keyless search provider and I close it myself. This is my last bump either way.