Spaces:
Sleeping
Sleeping
Upload 5 files
Browse files- aworld/memory/README.md +62 -0
- aworld/memory/__init__.py +2 -0
- aworld/memory/main.py +390 -0
- aworld/memory/models.py +197 -0
- aworld/memory/utils.py +33 -0
aworld/memory/README.md
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## multi-agents memory
|
2 |
+

|
3 |
+
|
4 |
+
### Short-Term Memory
|
5 |
+
|
6 |
+
Short-term memory (InMemory) is suitable for lightweight, temporary multi-agent memory scenarios. Data is only stored in memory, making it ideal for testing and small-scale experiments.
|
7 |
+
|
8 |
+
**Usage Example:**
|
9 |
+
|
10 |
+
```python
|
11 |
+
from aworld.core.memory import MemoryConfig, MemoryItem
|
12 |
+
from aworld.memory.main import MemoryFactory
|
13 |
+
|
14 |
+
# Create InMemory config
|
15 |
+
memory_config = MemoryConfig(provider="inmemory", enable_summary=False)
|
16 |
+
# Initialize Memory
|
17 |
+
memory = MemoryFactory.from_config(memory_config)
|
18 |
+
|
19 |
+
# Add a memory item
|
20 |
+
memory.add(MemoryItem(content="Hello, world!", metadata={"user_id": "u1"}, tags=["greeting"]))
|
21 |
+
|
22 |
+
# Get all memory items
|
23 |
+
all_memories = memory.get_all()
|
24 |
+
for item in all_memories:
|
25 |
+
print(item.content)
|
26 |
+
```
|
27 |
+
|
28 |
+
### Long-Term Memory
|
29 |
+
|
30 |
+
Long-term memory (Mem0) is suitable for persistent, vectorized retrieval and summarization in multi-agent scenarios. It supports LLM-based summarization and vector storage.
|
31 |
+
|
32 |
+
**Usage Example:**
|
33 |
+
|
34 |
+
```python
|
35 |
+
from aworld.core.memory import MemoryConfig, MemoryItem
|
36 |
+
from aworld.memory.main import MemoryFactory
|
37 |
+
|
38 |
+
# Create Mem0 config (requires mem0 and related dependencies)
|
39 |
+
memory_config = MemoryConfig(
|
40 |
+
provider="mem0",
|
41 |
+
enable_summary=True, # Enable summarization
|
42 |
+
summary_rounds=5, # Generate a summary every 5 rounds
|
43 |
+
embedder_provider="huggingface", # Embedding model provider
|
44 |
+
embedder_model="all-MiniLM-L6-v2", # Embedding model name
|
45 |
+
embedder_dims=384
|
46 |
+
)
|
47 |
+
# Initialize Memory
|
48 |
+
memory = MemoryFactory.from_config(memory_config)
|
49 |
+
|
50 |
+
# Add a memory item
|
51 |
+
memory.add(MemoryItem(content="The agent visited Hangzhou.", metadata={"user_id": "u1"}, tags=["travel"]))
|
52 |
+
|
53 |
+
# Get all memory items
|
54 |
+
all_memories = memory.get_all()
|
55 |
+
for item in all_memories:
|
56 |
+
print(item.content)
|
57 |
+
```
|
58 |
+
|
59 |
+
> Note: To use mem0, you must install `mem0` and `sentence-transformers` in advance, and configure the required LLM environment variables.
|
60 |
+
|
61 |
+
### CheckPoint
|
62 |
+
TODO
|
aworld/memory/__init__.py
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
aworld/memory/main.py
ADDED
@@ -0,0 +1,390 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding: utf-8
|
2 |
+
# Copyright (c) 2025 inclusionAI.
|
3 |
+
import abc
|
4 |
+
import asyncio
|
5 |
+
import json
|
6 |
+
import os
|
7 |
+
from typing import Optional
|
8 |
+
|
9 |
+
from aworld.config import ConfigDict
|
10 |
+
from aworld.core.memory import MemoryBase, MemoryItem, MemoryStore, MemoryConfig
|
11 |
+
from aworld.logs.util import logger
|
12 |
+
from aworld.models.llm import get_llm_model, acall_llm_model
|
13 |
+
|
14 |
+
|
15 |
+
class InMemoryMemoryStore(MemoryStore):
|
16 |
+
def __init__(self):
|
17 |
+
self.memory_items = []
|
18 |
+
|
19 |
+
def add(self, memory_item: MemoryItem):
|
20 |
+
self.memory_items.append(memory_item)
|
21 |
+
|
22 |
+
def get(self, memory_id) -> Optional[MemoryItem]:
|
23 |
+
return next((item for item in self.memory_items if item.id == memory_id), None)
|
24 |
+
|
25 |
+
def get_first(self, filters: dict = None) -> Optional[MemoryItem]:
|
26 |
+
"""Get the first memory item."""
|
27 |
+
filtered_items = self.get_all(filters)
|
28 |
+
if len(filtered_items) == 0:
|
29 |
+
return None
|
30 |
+
return filtered_items[0]
|
31 |
+
|
32 |
+
def total_rounds(self, filters: dict = None) -> int:
|
33 |
+
"""Get the total number of rounds."""
|
34 |
+
return len(self.get_all(filters))
|
35 |
+
|
36 |
+
def get_all(self, filters: dict = None) -> list[MemoryItem]:
|
37 |
+
"""Filter memory items based on filters."""
|
38 |
+
filtered_items = [item for item in self.memory_items if self._filter_memory_item(item, filters)]
|
39 |
+
return filtered_items
|
40 |
+
|
41 |
+
def _filter_memory_item(self, memory_item: MemoryItem, filters: dict = None) -> bool:
|
42 |
+
if memory_item.deleted:
|
43 |
+
return False
|
44 |
+
if filters is None:
|
45 |
+
return True
|
46 |
+
if filters.get('user_id') is not None:
|
47 |
+
if memory_item.metadata.get('user_id') is None:
|
48 |
+
return False
|
49 |
+
if memory_item.metadata.get('user_id') != filters['user_id']:
|
50 |
+
return False
|
51 |
+
if filters.get('agent_id') is not None:
|
52 |
+
if memory_item.metadata.get('agent_id') is None:
|
53 |
+
return False
|
54 |
+
if memory_item.metadata.get('agent_id') != filters['agent_id']:
|
55 |
+
return False
|
56 |
+
if filters.get('task_id') is not None:
|
57 |
+
if memory_item.metadata.get('task_id') is None:
|
58 |
+
return False
|
59 |
+
if memory_item.metadata.get('task_id') != filters['task_id']:
|
60 |
+
return False
|
61 |
+
if filters.get('session_id') is not None:
|
62 |
+
if memory_item.metadata.get('session_id') is None:
|
63 |
+
return False
|
64 |
+
if memory_item.metadata.get('session_id') != filters['session_id']:
|
65 |
+
return False
|
66 |
+
if filters.get('memory_type') is not None:
|
67 |
+
if memory_item.memory_type is None:
|
68 |
+
return False
|
69 |
+
if memory_item.memory_type != filters['memory_type']:
|
70 |
+
return False
|
71 |
+
return True
|
72 |
+
|
73 |
+
def get_last_n(self, last_rounds, filters: dict = None) -> list[MemoryItem]:
|
74 |
+
return self.memory_items[-last_rounds:] # Get the last n items
|
75 |
+
|
76 |
+
def update(self, memory_item: MemoryItem):
|
77 |
+
for index, item in enumerate(self.memory_items):
|
78 |
+
if item.id == memory_item.id:
|
79 |
+
self.memory_items[index] = memory_item # Update the item in the list
|
80 |
+
break
|
81 |
+
|
82 |
+
def delete(self, memory_id):
|
83 |
+
exists = self.get(memory_id)
|
84 |
+
if exists:
|
85 |
+
exists.deleted = True
|
86 |
+
|
87 |
+
def history(self, memory_id) -> list[MemoryItem] | None:
|
88 |
+
exists = self.get(memory_id)
|
89 |
+
if exists:
|
90 |
+
return exists.histories
|
91 |
+
return None
|
92 |
+
|
93 |
+
|
94 |
+
class MemoryFactory:
|
95 |
+
|
96 |
+
@classmethod
|
97 |
+
def from_config(cls, config: MemoryConfig) -> "MemoryBase":
|
98 |
+
"""
|
99 |
+
Initialize a Memory instance from a configuration dictionary.
|
100 |
+
|
101 |
+
Args:
|
102 |
+
config (dict): Configuration dictionary.
|
103 |
+
|
104 |
+
Returns:
|
105 |
+
InMemoryStorageMemory: Memory instance.
|
106 |
+
"""
|
107 |
+
if config.provider == "inmemory":
|
108 |
+
return InMemoryStorageMemory(
|
109 |
+
memory_store=InMemoryMemoryStore(),
|
110 |
+
config=config,
|
111 |
+
enable_summary=config.enable_summary,
|
112 |
+
summary_rounds=config.summary_rounds
|
113 |
+
)
|
114 |
+
elif config.provider == "mem0":
|
115 |
+
from aworld.memory.mem0.mem0_memory import Mem0Memory
|
116 |
+
return Mem0Memory(
|
117 |
+
memory_store=InMemoryMemoryStore(),
|
118 |
+
config=config
|
119 |
+
)
|
120 |
+
else:
|
121 |
+
raise ValueError(f"Invalid memory store type: {config.get('memory_store')}")
|
122 |
+
|
123 |
+
|
124 |
+
class Memory(MemoryBase):
|
125 |
+
__metaclass__ = abc.ABCMeta
|
126 |
+
|
127 |
+
def __init__(self, memory_store: MemoryStore, config: MemoryConfig, **kwargs):
|
128 |
+
self.memory_store = memory_store
|
129 |
+
self.config = config
|
130 |
+
self._llm_instance = None
|
131 |
+
|
132 |
+
@property
|
133 |
+
def default_llm_instance(self):
|
134 |
+
def get_env(key: str, default_key: str, default_val: object=None):
|
135 |
+
return os.getenv(key) if os.getenv(key) else os.getenv(default_key, default_val)
|
136 |
+
|
137 |
+
if not self._llm_instance:
|
138 |
+
self._llm_instance = get_llm_model(conf=ConfigDict({
|
139 |
+
"llm_model_name": get_env("MEM_LLM_MODEL_NAME", "LLM_MODEL_NAME"),
|
140 |
+
"llm_api_key": get_env("MEM_LLM_API_KEY", "LLM_MODEL_NAME") ,
|
141 |
+
"llm_base_url": get_env("MEM_LLM_BASE_URL", 'LLM_BASE_URL'),
|
142 |
+
"temperature": get_env("MEM_LLM_TEMPERATURE", "MEM_LLM_TEMPERATURE", 1.0),
|
143 |
+
"streaming": 'False'
|
144 |
+
}))
|
145 |
+
return self._llm_instance
|
146 |
+
|
147 |
+
def _build_history_context(self, messages) -> str:
|
148 |
+
"""Build the history context string from a list of messages.
|
149 |
+
|
150 |
+
Args:
|
151 |
+
messages: List of message objects with 'role', 'content', and optional 'tool_calls'.
|
152 |
+
Returns:
|
153 |
+
Concatenated context string.
|
154 |
+
"""
|
155 |
+
history_context = ""
|
156 |
+
for item in messages:
|
157 |
+
history_context += (f"\n\n{item['role']}: {item['content']}, "
|
158 |
+
f"{'tool_calls:' + json.dumps(item['tool_calls']) if 'tool_calls' in item and item['tool_calls'] else ''}")
|
159 |
+
return history_context
|
160 |
+
|
161 |
+
async def _call_llm_summary(self, summary_messages: list) -> str:
|
162 |
+
"""Call LLM to generate summary and log the process.
|
163 |
+
|
164 |
+
Args:
|
165 |
+
summary_messages: List of messages to send to LLM.
|
166 |
+
Returns:
|
167 |
+
Summary content string.
|
168 |
+
"""
|
169 |
+
logger.info(f"🤔 [Summary] Creating summary memory, history messages: {summary_messages}")
|
170 |
+
llm_response = await acall_llm_model(
|
171 |
+
self.default_llm_instance,
|
172 |
+
messages=summary_messages,
|
173 |
+
stream=False
|
174 |
+
)
|
175 |
+
logger.info(f'🤔 [Summary] summary_content: result is {llm_response.content[:400] + "...truncated"} ')
|
176 |
+
return llm_response.content
|
177 |
+
|
178 |
+
def _get_parsed_history_messages(self, history_items: list[MemoryItem]) -> list[dict]:
|
179 |
+
"""Get and format history messages for summary.
|
180 |
+
|
181 |
+
Args:
|
182 |
+
history_items: list[MemoryItem]
|
183 |
+
Returns:
|
184 |
+
List of parsed message dicts
|
185 |
+
"""
|
186 |
+
parsed_messages = [
|
187 |
+
{
|
188 |
+
'role': message.metadata['role'],
|
189 |
+
'content': message.content,
|
190 |
+
'tool_calls': message.metadata.get('tool_calls') if message.metadata.get('tool_calls') else None
|
191 |
+
}
|
192 |
+
for message in history_items]
|
193 |
+
return parsed_messages
|
194 |
+
|
195 |
+
async def async_gen_multi_rounds_summary(self, to_be_summary: list[MemoryItem]) -> str:
|
196 |
+
logger.info(
|
197 |
+
f"🤔 [Summary] Creating summary memory, history messages")
|
198 |
+
if len(to_be_summary) == 0:
|
199 |
+
return ""
|
200 |
+
parsed_messages = self._get_parsed_history_messages(to_be_summary)
|
201 |
+
history_context = self._build_history_context(parsed_messages)
|
202 |
+
|
203 |
+
summary_messages = [
|
204 |
+
{"role": "user", "content": self.config.summary_prompt.format(context=history_context)}
|
205 |
+
]
|
206 |
+
|
207 |
+
return await self._call_llm_summary(summary_messages)
|
208 |
+
|
209 |
+
async def async_gen_summary(self, filters: dict, last_rounds: int) -> str:
|
210 |
+
"""A tool for summarizing the conversation history."""
|
211 |
+
|
212 |
+
logger.info(f"🤔 [Summary] Creating summary memory, history messages [filters -> {filters}, "
|
213 |
+
f"last_rounds -> {last_rounds}]")
|
214 |
+
history_items = self.memory_store.get_last_n(last_rounds, filters=filters)
|
215 |
+
if len(history_items) == 0:
|
216 |
+
return ""
|
217 |
+
parsed_messages = self._get_parsed_history_messages(history_items)
|
218 |
+
history_context = self._build_history_context(parsed_messages)
|
219 |
+
|
220 |
+
summary_messages = [
|
221 |
+
{"role": "user", "content": self.config.summary_prompt.format(context=history_context)}
|
222 |
+
]
|
223 |
+
|
224 |
+
return await self._call_llm_summary(summary_messages)
|
225 |
+
|
226 |
+
async def async_gen_cur_round_summary(self, to_be_summary: MemoryItem, filters: dict, last_rounds: int) -> str:
|
227 |
+
if self.config.enable_summary and len(to_be_summary.content) < self.config.summary_single_context_length:
|
228 |
+
return to_be_summary.content
|
229 |
+
|
230 |
+
logger.info(f"🤔 [Summary] Creating summary memory, history messages [filters -> {filters}, "
|
231 |
+
f"last_rounds -> {last_rounds}]: to be summary content is {to_be_summary.content}")
|
232 |
+
history_items = self.memory_store.get_last_n(last_rounds, filters=filters)
|
233 |
+
if len(history_items) == 0:
|
234 |
+
return ""
|
235 |
+
parsed_messages = self._get_parsed_history_messages(history_items)
|
236 |
+
|
237 |
+
# Append the to_be_summary
|
238 |
+
parsed_messages.append({
|
239 |
+
"role": to_be_summary.metadata['role'],
|
240 |
+
"content": f"{to_be_summary.content}",
|
241 |
+
'tool_call_id': to_be_summary.metadata['tool_call_id'],
|
242 |
+
})
|
243 |
+
history_context = self._build_history_context(parsed_messages)
|
244 |
+
|
245 |
+
summary_messages = [
|
246 |
+
{"role": "user", "content": self.config.summary_prompt.format(context=history_context)}
|
247 |
+
]
|
248 |
+
|
249 |
+
return await self._call_llm_summary(summary_messages)
|
250 |
+
|
251 |
+
def search(self, query, limit=100, filters=None) -> Optional[list[MemoryItem]]:
|
252 |
+
pass
|
253 |
+
|
254 |
+
|
255 |
+
class InMemoryStorageMemory(Memory):
|
256 |
+
def __init__(self, memory_store: MemoryStore, config: MemoryConfig, enable_summary: bool = True, **kwargs):
|
257 |
+
super().__init__(memory_store=memory_store, config=config)
|
258 |
+
self.summary = {}
|
259 |
+
self.summary_rounds = self.config.summary_rounds
|
260 |
+
self.enable_summary = self.config.enable_summary
|
261 |
+
|
262 |
+
def add(self, memory_item: MemoryItem, filters: dict = None):
|
263 |
+
self.memory_store.add(memory_item)
|
264 |
+
|
265 |
+
# Check if we need to create or update summary
|
266 |
+
if self.enable_summary:
|
267 |
+
total_rounds = len(self.memory_store.get_all())
|
268 |
+
if total_rounds > self.summary_rounds:
|
269 |
+
self._create_or_update_summary(total_rounds)
|
270 |
+
|
271 |
+
def _create_or_update_summary(self, total_rounds: int):
|
272 |
+
"""Create or update summary based on current total rounds.
|
273 |
+
|
274 |
+
Args:
|
275 |
+
total_rounds (int): Total number of rounds.
|
276 |
+
"""
|
277 |
+
summary_index = int(total_rounds / self.summary_rounds)
|
278 |
+
start = (summary_index - 1) * self.summary_rounds
|
279 |
+
end = total_rounds - self.summary_rounds
|
280 |
+
|
281 |
+
# Ensure we have valid start and end indices
|
282 |
+
start = max(0, start)
|
283 |
+
end = max(start, end)
|
284 |
+
|
285 |
+
# Get the memory items to summarize
|
286 |
+
items_to_summarize = self.memory_store.get_all()[start:end + 1]
|
287 |
+
print(f"{total_rounds}start: {start}, end: {end},")
|
288 |
+
|
289 |
+
# Create summary content
|
290 |
+
summary_content = self._summarize_items(items_to_summarize, summary_index)
|
291 |
+
|
292 |
+
# Create the range key
|
293 |
+
range_key = f"{start}_{end}"
|
294 |
+
|
295 |
+
# Check if summary for this range already exists
|
296 |
+
if range_key in self.summary:
|
297 |
+
# Update existing summary
|
298 |
+
self.summary[range_key].content = summary_content
|
299 |
+
self.summary[range_key].updated_at = None # This will update the timestamp
|
300 |
+
else:
|
301 |
+
# Create new summary
|
302 |
+
summary_item = MemoryItem(
|
303 |
+
content=summary_content,
|
304 |
+
metadata={
|
305 |
+
"summary_index": summary_index,
|
306 |
+
"start_round": start,
|
307 |
+
"end_round": end,
|
308 |
+
"role": "system"
|
309 |
+
},
|
310 |
+
tags=["summary"]
|
311 |
+
)
|
312 |
+
self.summary[range_key] = summary_item
|
313 |
+
|
314 |
+
def _summarize_items(self, items: list[MemoryItem], summary_index: int) -> str:
|
315 |
+
"""Summarize a list of memory items.
|
316 |
+
|
317 |
+
Args:
|
318 |
+
items (list[MemoryItem]): List of memory items to summarize.
|
319 |
+
summary_index (int): Summary index.
|
320 |
+
|
321 |
+
Returns:
|
322 |
+
str: Summary content.
|
323 |
+
"""
|
324 |
+
# This is a placeholder. In a real implementation, you might use an LLM or other method
|
325 |
+
# to create a meaningful summary of the content
|
326 |
+
return asyncio.run(self.async_gen_multi_rounds_summary(items))
|
327 |
+
|
328 |
+
def update(self, memory_item: MemoryItem):
|
329 |
+
self.memory_store.update(memory_item)
|
330 |
+
|
331 |
+
def delete(self, memory_id):
|
332 |
+
self.memory_store.delete(memory_id)
|
333 |
+
|
334 |
+
def get(self, memory_id) -> Optional[MemoryItem]:
|
335 |
+
return self.memory_store.get(memory_id)
|
336 |
+
|
337 |
+
def get_all(self, filters: dict = None) -> list[MemoryItem]:
|
338 |
+
return self.memory_store.get_all()
|
339 |
+
|
340 |
+
def get_last_n(self, last_rounds, add_first_message=True, filters: dict = None) -> list[MemoryItem]:
|
341 |
+
"""Get last n memories.
|
342 |
+
|
343 |
+
Args:
|
344 |
+
last_rounds (int): Number of memories to retrieve.
|
345 |
+
add_first_message (bool):
|
346 |
+
|
347 |
+
Returns:
|
348 |
+
list[MemoryItem]: List of latest memories.
|
349 |
+
"""
|
350 |
+
memory_items = self.memory_store.get_last_n(last_rounds)
|
351 |
+
while len(memory_items) > 0 and memory_items[0].metadata and "tool_call_id" in memory_items[0].metadata and \
|
352 |
+
memory_items[0].metadata["tool_call_id"]:
|
353 |
+
last_rounds = last_rounds + 1
|
354 |
+
memory_items = self.memory_store.get_last_n(last_rounds)
|
355 |
+
|
356 |
+
# If summary is disabled or no summaries exist, return just the last_n_items
|
357 |
+
if not self.enable_summary or not self.summary:
|
358 |
+
return memory_items
|
359 |
+
|
360 |
+
# Calculate the range for relevant summaries
|
361 |
+
all_items = self.memory_store.get_all()
|
362 |
+
total_items = len(all_items)
|
363 |
+
end_index = total_items - last_rounds
|
364 |
+
|
365 |
+
# Get complete summaries
|
366 |
+
result = []
|
367 |
+
complete_summary_count = end_index // self.summary_rounds
|
368 |
+
|
369 |
+
# Get complete summaries
|
370 |
+
for i in range(complete_summary_count):
|
371 |
+
range_key = f"{i * self.summary_rounds}_{(i + 1) * self.summary_rounds - 1}"
|
372 |
+
if range_key in self.summary:
|
373 |
+
result.append(self.summary[range_key])
|
374 |
+
|
375 |
+
# Get the last incomplete summary if exists
|
376 |
+
remaining_items = end_index % self.summary_rounds
|
377 |
+
if remaining_items > 0:
|
378 |
+
start = complete_summary_count * self.summary_rounds
|
379 |
+
range_key = f"{start}_{end_index - 1}"
|
380 |
+
if range_key in self.summary:
|
381 |
+
result.append(self.summary[range_key])
|
382 |
+
|
383 |
+
# Add the last n items
|
384 |
+
result.extend(memory_items)
|
385 |
+
|
386 |
+
# Add first user input
|
387 |
+
if add_first_message and last_rounds < self.memory_store.total_rounds():
|
388 |
+
memory_items.insert(0, self.memory_store.get_first())
|
389 |
+
|
390 |
+
return result
|
aworld/memory/models.py
ADDED
@@ -0,0 +1,197 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pydantic import BaseModel, ConfigDict, Field
|
2 |
+
from aworld.core.memory import MemoryItem
|
3 |
+
from typing import Any, Dict, List, Optional, Literal
|
4 |
+
|
5 |
+
from aworld.models.model_response import ToolCall
|
6 |
+
|
7 |
+
class MessageMetadata(BaseModel):
|
8 |
+
"""
|
9 |
+
Metadata for memory messages, including user, session, task, and agent information.
|
10 |
+
Args:
|
11 |
+
user_id (str): The ID of the user.
|
12 |
+
session_id (str): The ID of the session.
|
13 |
+
task_id (str): The ID of the task.
|
14 |
+
agent_id (str): The ID of the agent.
|
15 |
+
"""
|
16 |
+
user_id: str = Field(description="The ID of the user")
|
17 |
+
session_id: str = Field(description="The ID of the session")
|
18 |
+
task_id: str = Field(description="The ID of the task")
|
19 |
+
agent_id: str = Field(description="The ID of the agent")
|
20 |
+
agent_name: str = Field(description="The name of the agent")
|
21 |
+
|
22 |
+
model_config = ConfigDict(extra="allow")
|
23 |
+
|
24 |
+
@property
|
25 |
+
def to_dict(self) -> Dict[str, Any]:
|
26 |
+
return self.model_dump()
|
27 |
+
|
28 |
+
class AgentExperienceItem(BaseModel):
|
29 |
+
skill: str = Field(description="The skill demonstrated in the experience")
|
30 |
+
actions: List[str] = Field(description="The actions taken by the agent")
|
31 |
+
|
32 |
+
class AgentExperience(MemoryItem):
|
33 |
+
"""
|
34 |
+
Represents an agent's experience, including skills and actions.
|
35 |
+
All custom attributes are stored in content and metadata.
|
36 |
+
Args:
|
37 |
+
agent_id (str): The ID of the agent.
|
38 |
+
skill (str): The skill demonstrated in the experience.
|
39 |
+
actions (List[str]): The actions taken by the agent.
|
40 |
+
metadata (Optional[Dict[str, Any]]): Additional metadata.
|
41 |
+
"""
|
42 |
+
def __init__(self, agent_id: str, skill: str, actions: List[str], metadata: Optional[Dict[str, Any]] = None) -> None:
|
43 |
+
meta = metadata.copy() if metadata else {}
|
44 |
+
meta['agent_id'] = agent_id
|
45 |
+
agent_experience = AgentExperienceItem(skill=skill, actions=actions)
|
46 |
+
super().__init__(content=agent_experience, metadata=meta, memory_type="agent_experience")
|
47 |
+
|
48 |
+
@property
|
49 |
+
def agent_id(self) -> str:
|
50 |
+
return self.metadata['agent_id']
|
51 |
+
|
52 |
+
@property
|
53 |
+
def skill(self) -> str:
|
54 |
+
return self.content.skill
|
55 |
+
|
56 |
+
@property
|
57 |
+
def actions(self) -> List[str]:
|
58 |
+
return self.content.actions
|
59 |
+
|
60 |
+
class UserProfileItem(BaseModel):
|
61 |
+
key: str = Field(description="The key of the profile")
|
62 |
+
value: Any = Field(description="The value of the profile")
|
63 |
+
|
64 |
+
class UserProfile(MemoryItem):
|
65 |
+
"""
|
66 |
+
Represents a user profile key-value pair.
|
67 |
+
All custom attributes are stored in content and metadata.
|
68 |
+
Args:
|
69 |
+
user_id (str): The ID of the user.
|
70 |
+
key (str): The profile key.
|
71 |
+
value (Any): The profile value.
|
72 |
+
metadata (Optional[Dict[str, Any]]): Additional metadata.
|
73 |
+
"""
|
74 |
+
def __init__(self, user_id: str, key: str, value: Any, metadata: Optional[Dict[str, Any]] = None) -> None:
|
75 |
+
meta = metadata.copy() if metadata else {}
|
76 |
+
meta['user_id'] = user_id
|
77 |
+
user_profile = UserProfileItem(key=key, value=value)
|
78 |
+
super().__init__(content=user_profile, metadata=meta, memory_type="user_profile")
|
79 |
+
|
80 |
+
@property
|
81 |
+
def user_id(self) -> str:
|
82 |
+
return self.metadata['user_id']
|
83 |
+
|
84 |
+
@property
|
85 |
+
def key(self) -> str:
|
86 |
+
return self.content.key
|
87 |
+
|
88 |
+
@property
|
89 |
+
def value(self) -> Any:
|
90 |
+
return self.content.value
|
91 |
+
|
92 |
+
class MemoryMessage(MemoryItem):
|
93 |
+
"""
|
94 |
+
Represents a memory message with role, user, session, task, and agent information.
|
95 |
+
Args:
|
96 |
+
role (str): The role of the message sender.
|
97 |
+
metadata (MessageMetadata): Metadata object containing user, session, task, and agent IDs.
|
98 |
+
content (Optional[Any]): Content of the message.
|
99 |
+
"""
|
100 |
+
def __init__(self, role: str, metadata: MessageMetadata, content: Optional[Any] = None) -> None:
|
101 |
+
meta = metadata.to_dict
|
102 |
+
meta['role'] = role
|
103 |
+
super().__init__(content=content, metadata=meta, memory_type="message")
|
104 |
+
|
105 |
+
@property
|
106 |
+
def role(self) -> str:
|
107 |
+
return self.metadata['role']
|
108 |
+
|
109 |
+
@property
|
110 |
+
def user_id(self) -> str:
|
111 |
+
return self.metadata['user_id']
|
112 |
+
|
113 |
+
@property
|
114 |
+
def session_id(self) -> str:
|
115 |
+
return self.metadata['session_id']
|
116 |
+
|
117 |
+
@property
|
118 |
+
def task_id(self) -> str:
|
119 |
+
return self.metadata['task_id']
|
120 |
+
|
121 |
+
@property
|
122 |
+
def agent_id(self) -> str:
|
123 |
+
return self.metadata['agent_id']
|
124 |
+
|
125 |
+
class SystemMessage(MemoryMessage):
|
126 |
+
"""
|
127 |
+
Represents a system message with role and content.
|
128 |
+
Args:
|
129 |
+
metadata (MessageMetadata): Metadata object containing user, session, task, and agent IDs.
|
130 |
+
content (str): The content of the message.
|
131 |
+
"""
|
132 |
+
def __init__(self, content: str, metadata: MessageMetadata) -> None:
|
133 |
+
super().__init__(role="system", metadata=metadata, content=content)
|
134 |
+
|
135 |
+
@property
|
136 |
+
def content(self) -> str:
|
137 |
+
return self._content
|
138 |
+
|
139 |
+
class HumanMessage(MemoryMessage):
|
140 |
+
"""
|
141 |
+
Represents a human message with role and content.
|
142 |
+
Args:
|
143 |
+
metadata (MessageMetadata): Metadata object containing user, session, task, and agent IDs.
|
144 |
+
content (str): The content of the message.
|
145 |
+
"""
|
146 |
+
def __init__(self, metadata: MessageMetadata, content: str) -> None:
|
147 |
+
super().__init__(role="human", metadata=metadata, content=content)
|
148 |
+
|
149 |
+
@property
|
150 |
+
def content(self) -> str:
|
151 |
+
return self._content
|
152 |
+
|
153 |
+
class AIMessage(MemoryMessage):
|
154 |
+
"""
|
155 |
+
Represents an AI message with role and content.
|
156 |
+
Args:
|
157 |
+
metadata (MessageMetadata): Metadata object containing user, session, task, and agent IDs.
|
158 |
+
content (str): The content of the message.
|
159 |
+
"""
|
160 |
+
def __init__(self, content: str, tool_calls: List[ToolCall], metadata: MessageMetadata) -> None:
|
161 |
+
meta = metadata.to_dict
|
162 |
+
meta['tool_calls'] = [tool_call.to_dict() for tool_call in tool_calls]
|
163 |
+
super().__init__(role="assistant", metadata=MessageMetadata(**meta), content=content)
|
164 |
+
|
165 |
+
@property
|
166 |
+
def content(self) -> str:
|
167 |
+
return self._content
|
168 |
+
|
169 |
+
@property
|
170 |
+
def tool_calls(self) -> List[ToolCall]:
|
171 |
+
return [ToolCall(**tool_call) for tool_call in self.metadata['tool_calls']]
|
172 |
+
|
173 |
+
class ToolMessage(MemoryMessage):
|
174 |
+
"""
|
175 |
+
Represents a tool message with role, content, tool_call_id, and status.
|
176 |
+
Args:
|
177 |
+
metadata (MessageMetadata): Metadata object containing user, session, task, and agent IDs.
|
178 |
+
tool_call_id (str): The ID of the tool call.
|
179 |
+
status (Literal["success", "error"]): The status of the tool call.
|
180 |
+
content (str): The content of the message.
|
181 |
+
"""
|
182 |
+
def __init__(self, tool_call_id: str, content: str, status: Literal["success", "error"] = "success", metadata: MessageMetadata = None) -> None:
|
183 |
+
metadata.tool_call_id = tool_call_id
|
184 |
+
metadata.status = status
|
185 |
+
super().__init__(role="tool", metadata=metadata, content=content)
|
186 |
+
|
187 |
+
@property
|
188 |
+
def tool_call_id(self) -> str:
|
189 |
+
return self.metadata['tool_call_id']
|
190 |
+
|
191 |
+
@property
|
192 |
+
def status(self) -> str:
|
193 |
+
return self.metadata['status']
|
194 |
+
|
195 |
+
@property
|
196 |
+
def content(self) -> str:
|
197 |
+
return self._content
|
aworld/memory/utils.py
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import tiktoken
|
2 |
+
from aworld.logs.util import logger
|
3 |
+
|
4 |
+
# TODO: merge to `models` package
|
5 |
+
|
6 |
+
MODEL_TO_ENCODING = {
|
7 |
+
"gpt-3.5-turbo": "cl100k_base",
|
8 |
+
"gpt-4": "cl100k_base",
|
9 |
+
"text-davinci-003": "p50k_base",
|
10 |
+
"text-embedding-ada-002": "cl100k_base",
|
11 |
+
"text-curie-001": "r50k_base",
|
12 |
+
"text-babbage-001": "r50k_base",
|
13 |
+
"text-ada-001": "r50k_base",
|
14 |
+
}
|
15 |
+
|
16 |
+
def get_encoding_for_model(model_name: str) -> tiktoken.Encoding:
|
17 |
+
"""
|
18 |
+
Automatically select the corresponding encoder based on the model name.
|
19 |
+
"""
|
20 |
+
encoding_name = MODEL_TO_ENCODING.get(model_name)
|
21 |
+
if encoding_name is None:
|
22 |
+
logger.warning(f"model '{model_name}' not found in mapping table.")
|
23 |
+
return "cl100k_base"
|
24 |
+
return encoding_name
|
25 |
+
|
26 |
+
def count_tokens(model_name: str, content: str):
|
27 |
+
encoding = tiktoken.get_encoding(get_encoding_for_model(model_name))
|
28 |
+
|
29 |
+
tokens = encoding.encode(content)
|
30 |
+
|
31 |
+
token_count = len(tokens)
|
32 |
+
|
33 |
+
return token_count
|