Add agentmemory LOCOMO benchmark comparison

Author: XelHaku

cmd/goncho-bench/locomo_backend_comparison.go

@@ -693,7 +693,7 @@ func setupNotesForBackend(name string) []string {
 	case "goncho-no-rank":
 		return []string{"Local deterministic no-ranking baseline in cmd/goncho-bench; uses the recency order before current Goncho ranking."}
 	case "agentmemory":
-		return []string{"External Python probe: scripts/bench_agentmemory_locomo.py --capability.", "Comparable when AGENTMEMORY_SOURCE_DIR points at PR #583 / commit 9b18a80c9d2839b025279978d3f4b5e1f9bc6e74 with npm dependencies installed.", "Adapter path uses standalone InMemoryKV fallback: memory_save external_id plus metadata.memory_id, then memory_smart_search. This validates stable IDs but is not the full running agentmemory server.", "If AGENTMEMORY_SOURCE_DIR is absent, agentmemory is marked not comparable."}
+		return []string{"External Python probe: scripts/bench_agentmemory_locomo.py --capability.", "Comparable when AGENTMEMORY_SOURCE_DIR points at https://github.com/rohitg00/agentmemory PR #583 / commit 9b18a80c9d2839b025279978d3f4b5e1f9bc6e74 with npm dependencies installed.", "Adapter path uses standalone InMemoryKV fallback: memory_save external_id plus metadata.memory_id, then memory_smart_search. This validates stable IDs but is not the full running agentmemory server.", "If AGENTMEMORY_SOURCE_DIR is absent, agentmemory is marked not comparable."}
 	case "mem0":
 		return []string{"External Python probe: scripts/bench_mem0_locomo.py --capability.", "Exact package version used in this run: none; backend marked not comparable before scoring.", "Install candidate: pip install mem0ai, with local/vector dependencies configured by upstream mem0 docs.", "Current status: not comparable in this harness until search results return caller-supplied memory_id unchanged without LLM answer scoring."}
 	default:
@@ -878,7 +878,7 @@ func writeLocomoBackendComparisonMarkdown(path string, report locomoBackendCompa
 	}
 	b.WriteString("## Setup notes\n\n")
 	b.WriteString("- Goncho, Goncho no-rank, BM25, and SQLite FTS5 are local Go adapters with no hosted dependency.\n")
-	b.WriteString("- agentmemory probe: `python3 scripts/bench_agentmemory_locomo.py --capability`. Comparable when `AGENTMEMORY_SOURCE_DIR` points at PR #583 / commit `9b18a80c9d2839b025279978d3f4b5e1f9bc6e74` with npm dependencies installed. This adapter uses the standalone InMemoryKV fallback, not the full running agentmemory server.\n")
+	b.WriteString("- agentmemory probe: `python3 scripts/bench_agentmemory_locomo.py --capability`. Comparable when `AGENTMEMORY_SOURCE_DIR` points at `https://github.com/rohitg00/agentmemory` PR #583 / commit `9b18a80c9d2839b025279978d3f4b5e1f9bc6e74` with npm dependencies installed. This adapter uses the standalone InMemoryKV fallback, not the full running agentmemory server.\n")
 	b.WriteString("- mem0 probe: `python3 scripts/bench_mem0_locomo.py --capability`. Exact package version used here: none; backend is marked not comparable before scoring. Candidate install: `pip install mem0ai` plus upstream local vector-store dependencies. Comparable only after configured local retrieval can return caller-supplied `memory_id` without answer-generation scoring.\n")
 	b.WriteString("\n## Interpretation\n\nBackends marked not comparable are excluded from score claims until they implement the `MemoryBackend` contract and return the same stable `memory_id` values that were inserted. This keeps the arena fair and prevents answer-generation or LLM-judge effects from leaking into retrieval metrics.\n")
 	return os.WriteFile(path, []byte(b.String()), 0o644)

docs-site/src/content/docs/operators/runbook.md

@@ -201,8 +201,9 @@ Operator rules:
 Backend probe commands:
 
 ```sh
-AGENTMEMORY_SOURCE_DIR=/path/to/agentmemory-pr583 python3 scripts/bench_agentmemory_locomo.py --capability
-AGENTMEMORY_SOURCE_DIR=/path/to/agentmemory-pr583 python3 scripts/bench_agentmemory_locomo.py --smoke
+# source: https://github.com/rohitg00/agentmemory at PR #583 commit 9b18a80c9d2839b025279978d3f4b5e1f9bc6e74
+AGENTMEMORY_SOURCE_DIR=/path/to/agentmemory python3 scripts/bench_agentmemory_locomo.py --capability
+AGENTMEMORY_SOURCE_DIR=/path/to/agentmemory python3 scripts/bench_agentmemory_locomo.py --smoke
 python3 scripts/bench_mem0_locomo.py --capability
 python3 scripts/bench_mem0_locomo.py --smoke
 ```
@@ -214,7 +215,7 @@ Expected current status:
 | Goncho | comparable | Local adapter returns stable IDs. |
 | BM25 | comparable | Local lexical baseline returns stable IDs. |
 | SQLite FTS5 | comparable | Local FTS baseline returns stable IDs. |
-| agentmemory | comparable with PR source | Set `AGENTMEMORY_SOURCE_DIR` to PR #583 commit `9b18a80c9d2839b025279978d3f4b5e1f9bc6e74`. Stable IDs work; standalone fallback LOCOMO score is `0.0000`, and this is not the full running server. |
+| agentmemory | comparable with PR source | Set `AGENTMEMORY_SOURCE_DIR` to `https://github.com/rohitg00/agentmemory` at PR #583 commit `9b18a80c9d2839b025279978d3f4b5e1f9bc6e74`. Stable IDs work; standalone fallback LOCOMO score is `0.0000`, and this is not the full running server. |
 | mem0 | not comparable | Package is not installed locally; no stable-ID run exists. |
 
 Primary outputs:

docs/benchmarks/external-backend-adapters.md

@@ -70,15 +70,15 @@ LOCOMO contains duplicate and near-duplicate content, including repeated content
 | Goncho no-rank | local Go harness | yes | Native LOCOMO `memory_id` | Local deterministic no-ranking baseline that uses recency order before current Goncho ranking. |
 | BM25 | local Go harness | yes | Native LOCOMO `memory_id` | Local deterministic lexical baseline. |
 | SQLite FTS5 | local Go SQLite FTS5 | yes | Native LOCOMO `memory_id` column | Local deterministic lexical baseline. |
-| agentmemory | `@agentmemory/agentmemory 0.9.20`, PR #583 commit `9b18a80c9d2839b025279978d3f4b5e1f9bc6e74` | yes, standalone fallback | `memory_save.external_id` plus `metadata.memory_id` returned by `memory_smart_search` | Stable IDs work. LOCOMO full score is `0.0` for the standalone InMemoryKV fallback because it uses strict all-term substring matching; this is not the full running agentmemory server. |
+| agentmemory | `https://github.com/rohitg00/agentmemory`, `@agentmemory/agentmemory 0.9.20`, PR #583 commit `9b18a80c9d2839b025279978d3f4b5e1f9bc6e74` | yes, standalone fallback | `memory_save.external_id` plus `metadata.memory_id` returned by `memory_smart_search` | Stable IDs work. LOCOMO full score is `0.0` for the standalone InMemoryKV fallback because it uses strict all-term substring matching; this is not the full running agentmemory server. |
 | mem0 | Python `3.12.3`; package not installed locally | no | Not executed | `mem0`/`mem0ai` is not installed in this environment; no stable-ID run can be produced. |
 
 ## Setup commands
 
 agentmemory candidate setup:
 
 ```bash
-git clone --branch feature/stable-external-memory-ids https://github.com/XelHaku/agentmemory.git
+git clone https://github.com/rohitg00/agentmemory.git
 cd agentmemory
 git checkout 9b18a80c9d2839b025279978d3f4b5e1f9bc6e74
 npm install --legacy-peer-deps