Add knowl and agentmemory as memory methods

[dat999zx]

Adds two agentic-memory systems the paper does not currently evaluate, wired the same way as the existing memory agents.

Why these two

Both resolve conflicts at write time, which is the behaviour FactConsolidation isolates, and neither has an FC score anywhere today. agentmemory publishes LongMemEval-S retrieval recall — a different task on a different metric. knowl publishes nothing outside this harness.

Configs are copied, not chosen

Every value comes from Simple_rag_bm25: retrieve_num: 10, temperature: 0.7, input_length_limit: 10000000, buffer_length: 200 — the same values BM25, Zep, Cognee, HippoRAG-v2, RAPTOR, GraphRAG, Self-RAG and all four embedding baselines already use.

The reader path is shared between both new methods and reproduces _handle_bm25_rag exactly: trailing newline per item, Memory i: labels, instruction trailing the facts under the generic system template. Retrieval goes through _extract_retrieval_query as the RAG handlers do — skipping it is not a small error, since the ~200 tokens of task boilerplate are byte-identical across all 100 questions.

Nothing is installed into this venv

• knowl runs as a Node subprocess over a sentinel-framed stdio protocol. The framing matters: its embedding runtime and SQLite bindings both write to stdout freely, and an unframed protocol eventually swallows a stray log line and desynchronises into a plausible score rather than an error.
• agentmemory runs as its own service and is reached over REST on :3111.

Both need one env var / one running process, documented in the module docstrings.

Input normalisation

Both ingest an identical parsed fact list, one record per write, in context order. Feeding raw 4096-char chunks was considered and rejected: agentmemory stores one memory per remember call, so a chunk lands as a single record holding ~70 facts, which measures chunking rather than memory.

Results on this harness

FC-SH SubEM, gpt-4o-mini reader:

	6k	262k
knowl	95.0	89.0
agentmemory	83.0	79.0

Single runs at temperature: 0.7 — directional, not publication-grade. Raw per-question results and full method notes: https://github.com/dat999zx/knowl/blob/main/benchmarks/memoryagentbench/mab/FINDINGS.md

Disclosure: I maintain knowl. agentmemory is not mine; I ran it because no FC number existed and the comparison seemed worth having. Happy to drop the knowl half of this PR if a maintainer-run method is preferred.

One asymmetry noticed while matching the reader

_handle_mem0_agent places its memories in a system message and appends a Current Time: line to the user turn; the RAG family does neither. Both methods here follow the RAG family. Flagging it in case cross-family comparisons are meant to hold reader layout constant.

Builds on #22 (UTF-8 result writes, --force), which is still open.