Mem chen jiang

experiments/benchmarks/HLE/runner.py

@@ -52,6 +52,8 @@
 from src.envs.hle                    import HLEEnv
 from src.solver.single_agent         import SingleAgentSolver
 from src.memory.methods.empty        import EmptyMemory
+from src.memory                      import GenerativeMASMemory, VoyagerMASMemory, MemoryBankMASMemory
+from src.utils                       import EmbeddingFunc
 from src.reasoning                   import ReasoningIO
 from src.llm.model_caller            import ModelCaller
 from src.llm.token_tracker           import token_tracker
@@ -390,14 +392,44 @@ def run(cfg: dict, logger: logging.Logger) -> None:
     )
     reasoning = ReasoningIO(llm_model=solver_caller)
 
-    working_dir = mem_cfg.get("working_dir", str(output_dir / "memory_store"))
-    Path(working_dir).mkdir(parents=True, exist_ok=True)
-    memory = EmptyMemory(
-        namespace=mem_cfg.get("namespace", "hle_empty"),
-        global_config={"working_dir": str(ts_dir / "memory_store")},
-        llm_model=None,
-        embedding_func=None,
-    )
+    memory_store_dir = str(ts_dir / "memory_store")
+    Path(memory_store_dir).mkdir(parents=True, exist_ok=True)
+
+    memory_method = exp_cfg.get("memory_method", "empty")
+    if memory_method == "empty":
+        memory = EmptyMemory(
+            namespace=mem_cfg.get("namespace", "hle_empty"),
+            global_config={"working_dir": memory_store_dir},
+            llm_model=None,
+            embedding_func=None,
+        )
+    elif memory_method in ("generative", "voyager", "memorybank"):
+        memory_caller = ModelCaller(
+            model=model_cfg["solver"],
+            role="memory",
+            base_url=base_url,
+        )
+        embedding_func = EmbeddingFunc()
+        mem_global_config = {
+            "working_dir":     memory_store_dir,
+            "successful_topk": mem_cfg.get("successful_topk", 1),
+            "failed_topk":     mem_cfg.get("failed_topk", 1),
+        }
+        mem_cls = {
+            "generative": GenerativeMASMemory,
+            "voyager":    VoyagerMASMemory,
+            "memorybank": MemoryBankMASMemory,
+        }[memory_method]
+        memory = mem_cls(
+            namespace=mem_cfg.get("namespace", f"hle_{memory_method}"),
+            global_config=mem_global_config,
+            llm_model=memory_caller,
+            embedding_func=embedding_func,
+        )
+        logger.info(f"Memory: {memory_method} (namespace={memory.namespace})")
+    else:
+        logger.error(f"Unknown memory_method: {memory_method}")
+        sys.exit(1)
 
     framework = exp_cfg.get("agent_framework", "single_agent")
     if framework == "single_agent":

experiments/configs/memory/generative.yaml

@@ -0,0 +1,8 @@
+experiment:
+  memory_method: generative
+
+memory_config:
+  namespace: hle_generative
+  working_dir: experiments/results/memory_store
+  successful_topk: 1
+  failed_topk: 1

experiments/configs/memory/memorybank.yaml

@@ -0,0 +1,8 @@
+experiment:
+  memory_method: memorybank
+
+memory_config:
+  namespace: hle_memorybank
+  working_dir: experiments/results/memory_store
+  successful_topk: 1
+  failed_topk: 1

experiments/configs/memory/voyager.yaml

@@ -0,0 +1,8 @@
+experiment:
+  memory_method: voyager
+
+memory_config:
+  namespace: hle_voyager
+  working_dir: experiments/results/memory_store
+  successful_topk: 1
+  failed_topk: 1

experiments/run_experiment.py

@@ -263,6 +263,10 @@ def main():
     except (FileNotFoundError, AttributeError) as e:
         logger.error(str(e))
         sys.exit(1)
+    except Exception as e:
+        logger.error(f"load_runner() 异常: {type(e).__name__}: {e}")
+        import traceback; traceback.print_exc()
+        sys.exit(1)
 
     runner.run(cfg, logger)
 

experiments/scripts/run_HLE_memory_test.sh

@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+# experiments/scripts/run_HLE_memory_test.sh
+#
+# 三种 Memory 方法的 HLE 适配测试脚本。
+#
+# 用法：
+#   # 正式跑（每种 50 题）
+#   bash experiments/scripts/run_HLE_memory_test.sh
+#
+#   # 冒烟测试（每种 2 题）
+#   bash experiments/scripts/run_HLE_memory_test.sh --smoke
+#
+#   # 只跑某一种（generative / voyager / memorybank）
+#   bash experiments/scripts/run_HLE_memory_test.sh --only generative
+
+set -euo pipefail
+unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY
+
+# ── API 配置 ──────────────────────────────────────────────────────────────────
+export OPENAI_API_KEY="sk-eb639510e766dc2868bc1974e678a055f6cba2bb351a74cb2696e46d24d360f3"
+export OPENAI_API_BASE="https://gmn.chuangzuoli.com"
+
+# ── 路径 ──────────────────────────────────────────────────────────────────────
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+cd "$REPO_ROOT"
+
+ENV_CFG="experiments/configs/envs/hle.yaml"
+SOLVER_CFG="experiments/configs/solver/single_agent.yaml"
+TOOL_CFG="experiments/configs/tool/default.yaml"
+
+# ── 参数解析 ──────────────────────────────────────────────────────────────────
+LIMIT=50
+ONLY=""
+for arg in "$@"; do
+    case $arg in
+        --smoke) LIMIT=2 ;;
+        --only)  shift; ONLY="$1" ;;
+    esac
+    shift 2>/dev/null || true
+done
+
+MODEL="gpt-5.2"
+OVERRIDES="evaluation.limit=$LIMIT evaluation.text_only=true model.solver=$MODEL model.judge=$MODEL model.base_url=https://gmn.chuangzuoli.com"
+
+# ── 运行函数 ──────────────────────────────────────────────────────────────────
+run_memory() {
+    local method=$1
+    local mem_cfg="experiments/configs/memory/${method}.yaml"
+
+    echo ""
+    echo "============================================================"
+    echo "  Memory: $method  |  Limit: $LIMIT  |  Model: $MODEL"
+    echo "============================================================"
+    echo ""
+
+    python experiments/run_experiment.py \
+        --env    "$ENV_CFG" \
+        --solver "$SOLVER_CFG" \
+        --tool   "$TOOL_CFG" \
+        --memory "$mem_cfg" \
+        --override $OVERRIDES
+}
+
+# ── 执行 ──────────────────────────────────────────────────────────────────────
+METHODS=("generative" "voyager" "memorybank")
+
+if [ -n "$ONLY" ]; then
+    run_memory "$ONLY"
+else
+    for method in "${METHODS[@]}"; do
+        run_memory "$method"
+    done
+fi
+
+echo ""
+echo "All experiments completed."

src/memory/__init__.py

@@ -2,20 +2,20 @@
 from .base import MemoryBase
 from .methods import (
     EmptyMemory,
-#     GenerativeMASMemory,
-#     VoyagerMASMemory,
-#     MemoryBankMASMemory,
+    GenerativeMASMemory,
+    VoyagerMASMemory,
+    MemoryBankMASMemory,
 #     ChatDevMASMemory,
 #     MetaGPTMASMemory,
 #     GMemory,
 )
 
 __all__ = [
-#     "MemoryBase",
+    "MemoryBase",
     "EmptyMemory",
-#    "GenerativeMASMemory",
-#    "VoyagerMASMemory",
-#    "MemoryBankMASMemory",
+    "GenerativeMASMemory",
+    "VoyagerMASMemory",
+    "MemoryBankMASMemory",
 #    "ChatDevMASMemory",
 #    "MetaGPTMASMemory",
 #    "GMemory",

src/memory/methods/__init__.py

@@ -1,17 +1,17 @@
 # src/memory/methods/__init__.py
 from .empty import EmptyMemory
-# from .generative import GenerativeMASMemory
-# from .voyager import VoyagerMASMemory
-# from .memorybank import MemoryBankMASMemory
+from .generative import GenerativeMASMemory
+from .voyager import VoyagerMASMemory
+from .memorybank import MemoryBankMASMemory
 # from .chatdev import ChatDevMASMemory
 # from .metagpt import MetaGPTMASMemory
 # from .GMemory import GMemory
 
 __all__ = [
     "EmptyMemory",
-    # "GenerativeMASMemory",
-    # "VoyagerMASMemory",
-    # "MemoryBankMASMemory",
+    "GenerativeMASMemory",
+    "VoyagerMASMemory",
+    "MemoryBankMASMemory",
     # "ChatDevMASMemory",
     # "MetaGPTMASMemory",
     # "GMemory",