Fix mem0 search returning nothing, and the OSS extractor on transformers 5.x

[bjmeetsfo]

Three defects found by calling every public API end to end — all 39 MCP tools and all 11 mem0.Memory methods — against a real store and the real /v1 gateway ASGI app, rather than the stub gateway the existing tests use.

1. mem0.search() returned nothing, for every query

_reshape_search_results reads selected_refs. A ContextPack now serves its refs grouped — groups: [{type, n, items: [{text, ...}]}] — and emits no selected_refs key at all. So mem0's primary read API returned {"results": []} on a store that plainly had matching content (search_raw showed it).

Both shapes are accepted now. A group's type is carried into each item's metadata as ref_type, so callers keep the event/entity distinction the flat shape gave them.

2. mem0.search() results were not addressable

Fixing (1) exposed this. Group items carry no id — source_ref is deliberately classified as debug-only lineage and stripped from serving items — so results got synthetic ref-N-… ids. Feeding those back into get / update / delete, which is mem0's documented flow, failed with found: false, HTTP 500, and deleted: false.

Real ids are recovered from get_all by exact text match: one extra request, made only when a synthetic id is actually present. Items that match nothing keep the synthetic id on purpose — a derived entity ref ("preference: drink is matcha") is a projection of a memory, not an addressable memory, and inventing an id for it would only move the failure downstream.

3. The OSS causal-LM extractor could not run on transformers ≥ 5

apply_chat_template(..., return_tensors="pt") returns a BatchEncoding on transformers 5.x where it used to return a bare tensor. The result was passed positionally into generate(), which read .shape off a dict-like and raised a bare AttributeError with no indication of the cause — segment_provider=oss was simply unusable.

Both shapes are accepted now. Verified against Qwen2.5-7B-Instruct on transformers 5.15.0.

Why the tests missed all three

test_mem0_complete.py exercises the mem0 client against a stub BaseHTTPRequestHandler that returns hand-written selected_refs payloads, so it only ever tested request shaping and never saw a real pack. Nothing at all covered oss_model_memory_segments.

Verification

• 8 tests added. 4 of the mem0 tests and 2 of the extractor tests fail on the unfixed code, with the original errors.
• Full mem0 + extractor suite on this branch: 29 passing.
• API sweep: 53/53 (39 MCP tools + 11 mem0 methods + variants), against a real store and the real gateway.
• Correctness checks on retrieve/search beyond non-emptiness — grounding, per-user scoping in both directions, id resolution, limit, and agreement between search and search_raw: 13/13.

🤖 Generated with Claude Code