Add GoodMemory as a third backend (local HTTP bridge)

[hjqcan]

Adds GoodMemory v0.7.5 as a local HTTP backend next to Mem0 Cloud/OSS.

What is included

• benchmarks/common/goodmemory_client.py: an async adapter around the published goodmemory-client==0.1.0 package. The official client owns the HTTP wire contract, caller/scope headers, bearer auth, retries, and routing metadata.
• --backend goodmemory plus --goodmemory-host in the LoCoMo, LongMemEval, and BEAM runners.
• README setup using goodmemory-http-bridge --recommended; embeddings are optional.

Contract and evidence boundaries

• Deterministic fixture ingestion: every benchmark turn is imported as a verified fact with its timestamp preserved. GoodMemory 0.7.5's default bridge policy does not persist assistant assertions directly, so benchmark assistant turns are imported as verified fixture facts while their original role is preserved both in the indexed text and in metadataPatch.attributes.sourceRole.
• Run isolation: each project-derived user id creates a separate GoodMemory scope. The documented in-memory bridge is fresh per process, and each run should use a fresh --project-name.
• Observable routing: the adapter requests auto by default and reports strategy fallback instead of hiding a retrieval downgrade.
• Published recall limit: the 0.7.5 recall-context contract returns at most 12 selected items and has no caller-controlled item limit. The documented comparable configuration therefore uses --top-k 10 --top-k-cutoffs 10; requests above 12 emit an explicit warning.
• No invented relevance score: the bridge publishes ranked items but no numeric score. The adapter preserves bridge order and records 0.0 instead of synthesizing a number that could influence the retrieval judge.
• No benchmark claim: the smoke below verifies integration behavior only; it is not evidence of benchmark quality or uplift.

Validation

Environment: Python 3.12.13, Node v22.14.0, Bun 1.3.14, npm 10.9.2.

• python -m unittest discover -s tests -v — 6/6 passed.
• python -m py_compile for the adapter and all three runners.
• All three runner --help paths load and expose the GoodMemory backend under Python 3.12.
• python -m pip check — no broken requirements.
• git diff --check.
• Live smoke against the exact published goodmemory@0.7.5 package with bearer auth and an in-memory recommended bridge: two timestamped fixture facts accepted, a same-scope query recalled the relevant deployment-region fact, and a different scope returned no results. The bridge reported retrievalTier: preset-recommended; embeddings were intentionally disabled for this contract smoke.

No benchmark answerer, judge, paid model, or score was run as part of this validation.

[hjqcan]

Refreshed the GoodMemory adapter in commit 6e3b204 to preserve each runner's timestamp / observation_date as a UTC observation prefix before deterministic ingestion. This closes a temporal-fidelity gap affecting BEAM, LongMemEval, and LoCoMo.

Validation:

• python -m unittest discover -s tests -v (3/3)
• python -m py_compile for the adapter and all three runners
• live smoke against the current GoodMemory HTTP bridge: write with Unix timestamp, recall returns [Observed at 2023-05-08T00:00:00Z] ...

The PR remains ready for maintainer review.

[hjqcan]

Refreshed this integration for the published GoodMemory v0.7.0 in acf0d21. The benchmark adapter now delegates the wire protocol to the official goodmemory-client package instead of maintaining a second HTTP implementation, reducing the adapter from 298 to 202 lines while adding caller/scope correctness and observable recall routing. The setup no longer requires an embedding provider: goodmemory-http-bridge --recommended provides the local BM25/entity/RRF tier, with dense retrieval optional.

Validation is current: 5/5 unit tests, all three runners compile, git diff --check passes, and a live bearer-authenticated write/recall smoke succeeded against npm goodmemory@0.7.0. GitHub reports the branch as mergeable; it is still awaiting maintainer review.

[hjqcan]

Current verification update: commit 5f70b8b refreshes the integration evidence and setup examples to the published GoodMemory v0.7.5.

Validation on Python 3.12.13, Node v22.14.0, Bun 1.3.14, and npm 10.9.2:

• 6/6 unit tests passed
• adapter and all three runners compile
• all three runner --help paths expose the GoodMemory backend
• pip check and git diff --check are clean
• live bearer-authenticated smoke against goodmemory@0.7.5: timestamped fixture ingestion, same-scope recall, and different-scope isolation

This is integration/runtime evidence only. No benchmark answerer, judge, paid model, or score was run. The PR is mergeable and still awaiting maintainer review.