Spaces:
Sleeping
Sleeping
File size: 13,716 Bytes
0b0338d 87b51e4 0b0338d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 | # server/memory_bank.py
"""
Episodic Memory Bank β v4.0
Cross-episode learning store for AI coding agents.
Every time an agent fails at a specific failure type, we store:
1. The failure pattern (what actions led to it)
2. The remediation hint (what should have been done)
3. A compact "lesson" that can be injected into future prompts
The memory grows across episodes. When a new episode starts:
- We retrieve the most relevant past lessons (by task similarity)
- We inject them as a "memory context" into the agent's system prompt
- This creates a real self-improvement loop
This is NOT implemented in any current agent framework as an
environment-side primitive. Devin, Copilot, etc. start fresh every run.
"""
from __future__ import annotations
import json
import time
import os
import hashlib
from typing import List, Dict, Any, Optional
from dataclasses import dataclass, field, asdict
@dataclass
class MemoryEntry:
"""One stored episode lesson."""
entry_id: str
episode_id: str
task: str
created_at: float
# Failure details
failure_type: str
failure_evidence: str
score: float
# Strategy used
strategy: str
action_sequence_hash: str # Compact fingerprint of the action pattern
# Lesson extracted
lesson_title: str
lesson_body: str # Full explanation of what went wrong
lesson_hint: str # Compact hint to inject into future prompts
lesson_plan: List[str] # Step-by-step corrective plan
# Retrieval metadata
relevance_tags: List[str] # Tags for retrieval (task1, write_file, read_before_write...)
times_retrieved: int = 0
times_helpful: int = 0 # Incremented when retry after this lesson improved score
def to_dict(self) -> dict:
return asdict(self)
@classmethod
def from_dict(cls, d: dict) -> "MemoryEntry":
return cls(**d)
@dataclass
class MemoryContext:
"""Injected memory context for a new episode."""
relevant_lessons: List[MemoryEntry]
system_prompt_injection: str # Full text to prepend to system prompt
user_context_injection: str # Full text to prepend to first user message
lessons_count: int
most_relevant_lesson: Optional[str]
class EpisodicMemoryBank:
"""
Persistent cross-episode memory bank.
Storage: JSON file on disk (or in-memory for Gradio sessions).
Each entry is a MemoryEntry with lesson + retrieval metadata.
Usage:
bank = EpisodicMemoryBank(persist_path="memory.json")
# After an episode:
bank.store(episode_result)
# Before next episode:
context = bank.retrieve(task="task1", max_lessons=3)
# Inject context.system_prompt_injection into agent
"""
MAX_ENTRIES = 50 # Keep last 50 lessons per task
def __init__(self, persist_path: Optional[str] = None):
self.persist_path = persist_path
self._entries: List[MemoryEntry] = []
if persist_path and os.path.exists(persist_path):
self._load()
def store(
self,
episode_id: str,
task: str,
failure_type: str,
failure_evidence: str,
score: float,
strategy: str,
trajectory_steps: List[dict],
improvement_plan: Optional[dict] = None,
) -> MemoryEntry:
"""Store a lesson from a completed episode."""
# Build action fingerprint
actions = [s.get("action_type", "?") for s in trajectory_steps]
seq_str = "β".join(actions[:12])
seq_hash = hashlib.md5(seq_str.encode()).hexdigest()[:8]
# Relevance tags for retrieval
tags = [task, failure_type, strategy]
if "read_file" in actions:
tags.append("read_file")
if "write_file" in actions:
tags.append("write_file")
if "run_tests" not in actions:
tags.append("no_verification")
if len(actions) <= 3:
tags.append("too_short")
# Extract lesson from improvement plan or failure type
if improvement_plan:
lesson_title = improvement_plan.get("failure_type", failure_type)
lesson_body = improvement_plan.get("what_went_wrong", "Agent failed.")
lesson_hint = improvement_plan.get("system_prompt_addon", "")
lesson_plan = improvement_plan.get("step_by_step_plan", [])
else:
lesson_title, lesson_body, lesson_hint, lesson_plan = self._default_lesson(
failure_type, score, strategy
)
entry = MemoryEntry(
entry_id=f"{task}_{seq_hash}_{int(time.time())}",
episode_id=episode_id,
task=task,
created_at=time.time(),
failure_type=failure_type,
failure_evidence=failure_evidence[:200],
score=score,
strategy=strategy,
action_sequence_hash=seq_hash,
lesson_title=lesson_title,
lesson_body=lesson_body,
lesson_hint=lesson_hint,
lesson_plan=lesson_plan,
relevance_tags=tags,
times_retrieved=0,
times_helpful=0,
)
self._entries.append(entry)
self._trim()
if self.persist_path:
self._save()
return entry
def retrieve(
self,
task: str,
failure_type: Optional[str] = None,
strategy: Optional[str] = None,
max_lessons: int = 3,
) -> MemoryContext:
"""Retrieve relevant lessons for an upcoming episode."""
if not self._entries:
return MemoryContext(
relevant_lessons=[],
system_prompt_injection="",
user_context_injection="",
lessons_count=0,
most_relevant_lesson=None,
)
# Score each entry by relevance
scored: List[tuple[float, MemoryEntry]] = []
for e in self._entries:
score = 0.0
if e.task == task:
score += 3.0
elif task in e.relevance_tags:
score += 2.0
if failure_type and e.failure_type == failure_type:
score += 2.0
if strategy and e.strategy == strategy:
score += 1.0
# Penalize already-retrieved lessons slightly (freshness)
score -= e.times_retrieved * 0.1
# Boost low-score lessons (more informative failures)
score += max(0, 0.5 - e.score)
scored.append((score, e))
scored.sort(key=lambda x: -x[0])
relevant = [e for _, e in scored[:max_lessons]]
# Mark as retrieved
for e in relevant:
e.times_retrieved += 1
if not relevant:
return MemoryContext(
relevant_lessons=[],
system_prompt_injection="",
user_context_injection="",
lessons_count=0,
most_relevant_lesson=None,
)
# Build injection text
sys_lines = [
"π§ AGENT MEMORY β LESSONS FROM PAST EPISODES",
"=" * 50,
"You have made these mistakes before. Do NOT repeat them.",
"",
]
for i, e in enumerate(relevant, 1):
sys_lines += [
f"[Lesson {i}] Task: {e.task} | Failure: {e.failure_type} | Score was: {e.score:.2f}",
f"What went wrong: {e.lesson_body}",
f"IMPORTANT: {e.lesson_hint}" if e.lesson_hint else "",
"",
]
sys_lines.append("=" * 50)
system_injection = "\n".join(l for l in sys_lines if l is not None)
user_lines = [
"[MEMORY CONTEXT β Read before you act]",
]
for i, e in enumerate(relevant, 1):
user_lines.append(f"Past lesson {i}: {e.lesson_title}")
if e.lesson_plan:
user_lines.append("Correct approach:")
user_lines.extend(f" {step}" for step in e.lesson_plan[:4])
user_injection = "\n".join(user_lines)
return MemoryContext(
relevant_lessons=relevant,
system_prompt_injection=system_injection,
user_context_injection=user_injection,
lessons_count=len(relevant),
most_relevant_lesson=relevant[0].lesson_title if relevant else None,
)
def get_all_entries(self) -> List[dict]:
return [e.to_dict() for e in self._entries]
def get_stats(self) -> dict:
if not self._entries:
return {"total_entries": 0, "tasks": {}}
from collections import Counter
failure_counts = Counter(e.failure_type for e in self._entries)
task_counts = Counter(e.task for e in self._entries)
avg_score = sum(e.score for e in self._entries) / len(self._entries)
return {
"total_entries": len(self._entries),
"average_score_of_stored_episodes": round(avg_score, 3),
"failure_breakdown": dict(failure_counts.most_common()),
"tasks": dict(task_counts),
"most_helpful_lesson": max(self._entries, key=lambda e: e.times_helpful).lesson_title
if any(e.times_helpful > 0 for e in self._entries) else None,
}
def mark_helpful(self, episode_id: str):
"""Call this when a retry with a lesson improved the score."""
for e in self._entries:
if e.episode_id == episode_id:
e.times_helpful += 1
if self.persist_path:
self._save()
def clear(self, task: Optional[str] = None):
if task:
self._entries = [e for e in self._entries if e.task != task]
else:
self._entries = []
if self.persist_path:
self._save()
# ββ Persistence βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _save(self):
tpath = getattr(self, "persist_path", None)
if not tpath:
return
try:
with open(tpath, "w") as f:
json.dump([e.to_dict() for e in self._entries], f, indent=2)
except PermissionError:
import tempfile
fallback = os.path.join(tempfile.gettempdir(), "agent_memory.json")
if self.persist_path != fallback:
self.persist_path = fallback
self._save()
except Exception as e:
print(f"[MemoryBank] Warning: Failed to save to {self.persist_path}: {e}")
def _load(self):
try:
with open(self.persist_path, "r") as f:
data = json.load(f)
self._entries = [MemoryEntry.from_dict(d) for d in data]
except Exception:
self._entries = []
def _trim(self):
"""Keep at most MAX_ENTRIES, dropping oldest low-score entries first."""
if len(self._entries) <= self.MAX_ENTRIES:
return
# Sort by: useful first, then by recency
self._entries.sort(
key=lambda e: (
-e.times_helpful,
-e.times_retrieved,
e.created_at,
),
reverse=True,
)
self._entries = self._entries[:self.MAX_ENTRIES]
def _default_lesson(
self, failure_type: str, score: float, strategy: str
) -> tuple[str, str, str, List[str]]:
lessons = {
"NEVER_TESTED": (
"Submitted without verification",
"Agent submitted code without running tests. No confidence in correctness.",
"CRITICAL: Run run_tests after EVERY write_file. Never submit without test verification.",
["1. Write fix", "2. run_tests to check", "3. If passing β submit", "4. If failing β re-read and fix"],
),
"BLIND_WRITE": (
"Wrote without reading",
"Agent wrote to a file without reading it first. Blind writes introduce new bugs.",
"NEVER use write_file before read_file on the same path.",
["1. read_file first", "2. Understand existing code", "3. Then write minimal fix"],
),
"WRONG_FILE_NAVIGATION": (
"Navigated to wrong files",
"Agent read files unrelated to the bug. Wasted steps and missed root cause.",
"ALWAYS start with the failing test file. Its imports show you exactly where to go.",
["1. Read failing test", "2. Find its imports", "3. Navigate ONLY there"],
),
"LOOPING_BEHAVIOR": (
"Read same files repeatedly",
f"Agent looped reading the same files without progress. Score={score:.2f}.",
"Each file may be read AT MOST ONCE. Use search_code if confused.",
["1. Use search_code with function name", "2. Read matched file β once", "3. commit to fix"],
),
}
defaults = lessons.get(failure_type, (
f"{failure_type} failure",
f"Agent failed with type '{failure_type}', score={score:.2f}.",
"Read test β read source β fix β run_tests β submit.",
["1. read test", "2. read source", "3. write fix", "4. run_tests", "5. submit"],
))
return defaults
# Global singleton (shared across the Gradio session)
_GLOBAL_MEMORY = EpisodicMemoryBank(
persist_path=os.path.join(
os.path.dirname(__file__), "..", "agent_memory.json"
)
)
def get_global_memory() -> EpisodicMemoryBank:
return _GLOBAL_MEMORY
|