[lanbass869-cell]

Adds MemTether as a new memory method. Cross-client AI memory hub with file-level pointers, dual-timeline governance, conflict-aware exchange.

Benchmark numbers (updated 2026-10-20 — the original submission quoted a 2026-09-27 run):

• LongMemEval 500Q (full run, 2026-10-18): strict 62.4% (206/330), LLM-judge 68.8% (340/494). Per-type judge: knowledge-update 78.9%, ss-assistant 48%. multi-session judge 79.7% (106/133), vs mem0 60.2% on the same subset (+19.5pp).
• LoCoMo 1986Q (2026-10-14): token-F1 0.523, ranked 4th against 9 systems — ahead of Graphiti, mem0, SimpleMem, memU and Hindsight; behind PropMem, OpenClaw and FullContext. Per-category: multi-hop 0.556, open-domain 0.820, factual 0.357, temporal 0.264, inferential 0.206.

Both sets are reproducible; the repo's README documents the commands. Numbers that did not improve are stated as-is rather than cherry-picked.

Since the original submission, MemTether also gained:

• MCP surface: 10 tools (add/search/list/correct/retire/pin/as_of/timeline/profile/feedback) — agents can govern memory, not just search it.
• Layered recall (L0/L1/L2) — same hit at three depths; measured L0 costs 10% of full-text characters.
• Deterministic subject profiles (zero LLM, fully traceable to uids).
• 615 tests on the release repo (was 39) plus 125 on the internal hub repo, across a 9-job CI matrix (Linux/Windows/macOS × Python 3.10/3.12 + Docker).

Repo: https://github.com/MemTether/MemTether (Apache-2.0)
PyPI: https://pypi.org/project/memtether/ (0.1.0a83)
Docs: https://memtether.github.io/MemTether/

Related: awesome-agent-memory PR OpenDataBox/awesome-agent-memory#10

[lanbass869-cell]

Update: MemTether is now at v0.1.0a60 — 451 tests passing (up from 39 at initial submission), 9 CI jobs green, MCP Registry synced, Memory Court Playground live at https://memtether.github.io/MemTether/court/ . Still active and maintained — happy to adjust formatting to match list guidelines if needed.