fix(mem0): pass top_k to SDK in OSSBackend.get_all for accurate count

[luxuguang-leo]

Problem

OSSBackend.get_all() calls Memory.get_all(filters=filters) without passing top_k. The Mem0 SDK defaults top_k to 20 (source: mem0ai/mem0/memory/main.py), so the response is always truncated to at most 20 results.

Since OSSBackend computes count = len(all_results) from the raw SDK response, the reported count is wrong whenever there are more than 20 stored memories — it caps at 20.

Root Cause

In plugins/memory/mem0/_backend.py, OSSBackend.get_all():

def get_all(self, *, filters, page=1, page_size=100):
    response = self._memory.get_all(filters=filters)  # ← no top_k!
    all_results = _unwrap_results(response)
    total = len(all_results)  # ← always ≤ 20
    ...

The Mem0 SDK's Memory.get_all() signature is:

def get_all(self, *, filters=None, top_k: int = 20, show_expired=False):

The PlatformBackend does not have this problem — it uses MemoryClient.get_all(filters, page, page_size) which handles pagination server-side and returns an accurate count.

Fix

Pass top_k=10000 to self._memory.get_all() so the full result set is returned. OSSBackend already applies client-side pagination (page/page_size slicing) on the full set, so count = len(all_results) is correct again.

Testing

4 new tests in TestOSSBackendGetAll:

Test	What it verifies
test_passes_top_k_to_sdk	top_k=10000 is forwarded to the SDK
test_count_reflects_total_not_sdk_default	50 memories → count=50 (not 20)
test_client_side_pagination	Page 2, size 10 returns items 10-19 with count=50
test_pagination_beyond_total_returns_empty	Out-of-range page returns [] with correct count

All 46 mem0 tests pass.

Fixes #52921

[luxuguang-leo]

Quick check-in. This PR's approach was independently confirmed today by two other contributors — #58323 (bvisible, running in production) and #58410 (nankingjing) were both triaged as duplicates of this one.

The check suites are stuck at action_required — CI needs a maintainer to approve the first run for an external fork. Is there anything else needed on this PR before that can happen? @alt-glitch

[luxuguang-leo]

This PR is obsoleted by commit 9a32272 (kshitijk4poor), which pruned get_all() from the ABC and all three backends — its only caller mem0_list was already removed by an earlier recall-tuning commit.

Related:

• fix(mem0): OSSBackend.get_all silently capped at 20 memories #58323 by @bvisible — same fix, independently confirmed
• fix(mem0): pass top_k to SDK in OSSBackend.get_all (salvage #52973) #58410 by @nankingjing — salvage duplicate
• feat(mem0): self-hosted dashboard backend + recall tuning (salvage #55614) #56943 — the salvage PR that led to the prune

Closing as no longer needed — the dead code path this PR fixed no longer exists on main.