fix: stop get_cache handing out the stored KV cache by reference

[jeojdi1]

Description

_concat_caches returns caches[0] unchanged when a single cache id is requested, so get_cache() hands the caller the stored KVCacheItem.memory object itself. The caller passes that cache to generate(), which appends to it in place — so the saved activation memory grows on every chat turn.

Measured across three turns, the stored cache length went 6 → 19 → 32 → 45. Nothing in the API suggests that retrieving a memory mutates it, and the multi-cache path already builds a fresh container, so only this early return leaks the reference.

This PR returns a new DynamicCache sharing the same tensors. The tensors are deliberately not cloned: generate() appends along the sequence axis rather than writing into existing rows, so a fresh container is sufficient to protect the stored item without paying to duplicate the cache.

Related Issue (Required): #2301

Type of change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

• Unit Test

test_get_cache_does_not_alias_stored_memory asserts the returned object is not the stored one, then simulates generate() by appending a step to the returned cache and asserts the stored length is unchanged.

Verified fails-before / passes-after on transformers 5.16.1:

# upstream
FAILED tests/memories/activation/test_kv.py::test_get_cache_does_not_alias_stored_memory
1 failed

# with this change
1 passed

Checklist

• I have performed a self-review of my own code
• I have commented my code in hard-to-understand areas
• I have added tests that prove my fix is effective
• I have created related documentation issue/PR in MemOS-Docs (not applicable)
• I have linked the issue to this PR (The stored activation DynamicCache is mutated in place by generation, so activation memory grows every turn #2301)
• I have mentioned the person who will review this PR

A note on the target branch

CONTRIBUTING.md says to open PRs against dev, but no dev branch exists — only main and dev-v2.0.28 … dev-v2.0.32. This is against main (185ebdb, "Dev v2.0.32"). Happy to retarget.

[Memtensor-AI]

🤖 Open Code Review

Target: PR #2303
Task: c9892125fee30ed9
Base: main
Head: fix/kv-cache-mutated-in-place-by-generate

🔍 OpenCodeReview found 2 issue(s) in this PR.

---

1. src/memos/memories/activation/kv.py (L92)

torch is imported but never referenced anywhere in _copy_cache. The # noqa: F401 suppressor confirms the linter already caught this. Unlike _concat_caches which actually calls torch.cat, this method performs only Python-level attribute copies. Remove the import.

💡 Suggested Change

Before:

        import torch  # noqa: F401  (kept local, matching this module's style)

After:

        # torch not needed here; tensors are shared by reference, not constructed

---

2. src/memos/memories/activation/kv.py (L103-L108)

dst.lazy_initialization(src.keys, src.values) is called conditionally, but dst.keys = src.keys and dst.values = src.values are then assigned unconditionally regardless of whether lazy_initialization ran. This has two problems:

1. If lazy_initialization only sets keys and values, the call is always redundant (the unconditional assignments below handle it).
2. If lazy_initialization does additional setup beyond setting those fields (e.g., allocating internal buffers or registering tensors), the unconditional overwrites on the next two lines may leave dst in an inconsistent internal state.

Consider one of:

• Remove the lazy_initialization call and rely solely on the unconditional assignments (correct for the shallow-copy intent documented in the docstring).
• Use an else branch so the assignments only happen when lazy_initialization was NOT called, keeping both paths mutually exclusive.

💡 Suggested Change

Before:

                if not getattr(dst, "is_initialized", True) and hasattr(
                    dst, "lazy_initialization"
                ):
                    dst.lazy_initialization(src.keys, src.values)
                dst.keys = src.keys
                dst.values = src.values

After:

                # Unconditional shallow assignment is sufficient for a
                # shallow-structure copy; lazy_initialization is not needed.
                dst.keys = src.keys
                dst.values = src.values

Generated by cloud-assistant via Open Code Review.

[Memtensor-AI]

⚠️ Automated Test Results: ENV ISSUE

The test environment encountered an issue that requires manual attention.

Details: Test collection failed because the torch module is not installed in the test environment, preventing the test file from being imported. [advisory, non-gating] AI-generated tests on branch test/auto-gen-c9892125fee30ed9-20260829013054: 36/36 passed — these do NOT affect the PR verdict; review the branch manually.
Branch: fix/kv-cache-mutated-in-place-by-generate