Skip to content

Agent Zero AI framework

Improve auto-memory extraction quality

Alessandroa0a815943a3b
diff --git a/plugins/_memory/README.md b/plugins/_memory/README.md
index 9b6d9617c..2f33a0b0b 100644
--- a/plugins/_memory/README.md
+++ b/plugins/_memory/README.md
@@ -15,6 +15,8 @@ This plugin stores memories and knowledge embeddings in a FAISS-backed vector da
- Loads configured knowledge directories into memory when a database is initialized.
- **Memory tools**
- Includes tools for saving, loading, deleting, forgetting, and behavior adjustment workflows.
+- **Automatic conversation memory**
+ - Stores durable preferences, project facts, and recurring constraints while filtering transient action-history fragments before insertion.
- **Dashboard APIs**
- Exposes search, delete, bulk delete, update, and subdirectory listing endpoints for the memory dashboard.
- **Scoped storage**
diff --git a/plugins/_memory/extensions/python/monologue_end/_50_memorize_fragments.py b/plugins/_memory/extensions/python/monologue_end/_50_memorize_fragments.py
index 2ab9fbf31..59bd71227 100644
--- a/plugins/_memory/extensions/python/monologue_end/_50_memorize_fragments.py
+++ b/plugins/_memory/extensions/python/monologue_end/_50_memorize_fragments.py
@@ -8,6 +8,7 @@ from helpers.defer import DeferredTask, THREAD_BACKGROUND
# Direct import - this extension lives inside the memory plugin
from plugins._memory.helpers.memory import Memory
+from plugins._memory.helpers.memory_quality import filter_auto_memory_fragments
from plugins._memory.tools.memory_load import DEFAULT_THRESHOLD as DEFAULT_MEMORY_THRESHOLD
@@ -104,9 +105,24 @@ class MemorizeMemories(Extension):
if not isinstance(memories, list) or len(memories) == 0:
log_item.update(heading="No useful information to memorize.")
return
- else:
- memories_txt = "\n\n".join([str(memory) for memory in memories]).strip()
- log_item.update(heading=f"{len(memories)} entries to memorize.", memories=memories_txt)
+
+ raw_memories_count = len(memories)
+ memories = filter_auto_memory_fragments(memories)
+ filtered_memories_count = raw_memories_count - len(memories)
+
+ if not memories:
+ log_item.update(
+ heading="No durable information to memorize.",
+ filtered_memories_count=filtered_memories_count,
+ )
+ return
+
+ memories_txt = "\n\n".join(memories).strip()
+ log_item.update(
+ heading=f"{len(memories)} durable entries to memorize.",
+ memories=memories_txt,
+ filtered_memories_count=filtered_memories_count,
+ )
# Process memories with intelligent consolidation
total_processed = 0
diff --git a/plugins/_memory/helpers/memory_quality.py b/plugins/_memory/helpers/memory_quality.py
new file mode 100644
index 000000000..82139dd1d
--- /dev/null
+++ b/plugins/_memory/helpers/memory_quality.py
@@ -0,0 +1,88 @@
+from __future__ import annotations
+
+import re
+from typing import Iterable
+
+
+_DURABLE_SUBJECT_RE = re.compile(
+ r"\b(user|project|repo|repository|workspace|runtime|service|server|"
+ r"plugin|agent zero|a0|profile|team|organization|company)\b",
+ re.IGNORECASE,
+)
+_DURABLE_RELATION_RE = re.compile(
+ r"\b(prefers?|requires?|uses?|runs?|lives?|located|configured|default|"
+ r"current(?:ly)?|must|should|always|never|constraint|path|endpoint|port)\b",
+ re.IGNORECASE,
+)
+_TRANSIENT_RE = re.compile(
+ r"(/tmp\b|\btmp/|\btemporary\b|\btemp\b|\bworkdir\b|"
+ r"\bcontainer-local\b|\bmachine-local\b|\bpersonal absolute path\b|"
+ r"\blocal endpoint\b|\blocal port\b|"
+ r"\btest marker\b|\bmemory test\b|\bcleanup token\b|\bthrowaway\b|"
+ r"\bsample file\b|\bdemo file\b|\bvalidation directory\b|\blive-test\b|"
+ r"\bone-off\b|\bsession-only\b)",
+ re.IGNORECASE,
+)
+_LOCAL_COORDINATE_RE = re.compile(
+ r"\b(localhost|127\.0\.0\.1|0\.0\.0\.0|::1)\b|"
+ r"https?://(?:localhost|127\.0\.0\.1|0\.0\.0\.0|\[::1\])(?::\d+)?|"
+ r"/(?:home|Users)/[^/\s]+/|"
+ r"\b[A-Za-z]:\\Users\\",
+ re.IGNORECASE,
+)
+_AGENT_ACTION_RE = re.compile(
+ r"\b(agent|assistant|codex|a0)\b.{0,80}\b("
+ r"created|implemented|built|ran|tested|verified|fixed|wrote|generated|"
+ r"saved|cleaned|inspected|synced|restarted|committed|staged|opened|"
+ r"reported|showed)\b",
+ re.IGNORECASE,
+)
+_USER_REQUEST_RE = re.compile(
+ r"\b(user|the user)\s+(asked|requested|wanted|told|prompted|instructed)\b",
+ re.IGNORECASE,
+)
+_COMMAND_HISTORY_RE = re.compile(
+ r"\b(ran|executed)\s+[`'\"]?[\w./-]+|\bterminal command\b|"
+ r"\bcommand output\b|\bexit code\b|\bstdout\b|\bstderr\b",
+ re.IGNORECASE,
+)
+_LOW_VALUE_RE = re.compile(
+ r"^(hello|hi|greeting|conversation start|task completed|file created|"
+ r"implementation completed|test passed|memory search)$",
+ re.IGNORECASE,
+)
+
+
+def normalize_memory_candidate(value: object) -> str:
+ return str(value).strip()
+
+
+def is_auto_fragment_worth_saving(text: str) -> bool:
+ """Return True only for durable facts suitable for automatic fragments."""
+
+ text = normalize_memory_candidate(text)
+ has_durable_relation = bool(_DURABLE_RELATION_RE.search(text))
+ if len(text) < 24:
+ return False
+ if _LOW_VALUE_RE.match(text):
+ return False
+ if _TRANSIENT_RE.search(text):
+ return False
+ if _LOCAL_COORDINATE_RE.search(text):
+ return False
+ if _COMMAND_HISTORY_RE.search(text):
+ return False
+ if _AGENT_ACTION_RE.search(text) and not has_durable_relation:
+ return False
+ if _USER_REQUEST_RE.search(text) and not has_durable_relation:
+ return False
+
+ return bool(_DURABLE_SUBJECT_RE.search(text) and has_durable_relation)
+
+
+def filter_auto_memory_fragments(memories: Iterable[object]) -> list[str]:
+ return [
+ text
+ for text in (normalize_memory_candidate(memory) for memory in memories)
+ if is_auto_fragment_worth_saving(text)
+ ]
diff --git a/plugins/_memory/prompts/fw.memory.hist_suc.sys.md b/plugins/_memory/prompts/fw.memory.hist_suc.sys.md
index 8739785e7..ddac2274b 100644
--- a/plugins/_memory/prompts/fw.memory.hist_suc.sys.md
+++ b/plugins/_memory/prompts/fw.memory.hist_suc.sys.md
@@ -20,4 +20,7 @@
# Rules
- Focus on important details like libraries used, code, encountered issues, error fixing etc.
-- Do not include simple solutions that don't require instructions to reproduce like file handling, web search etc.
\ No newline at end of file
+- Do not include simple solutions that don't require instructions to reproduce like file handling, web search etc.
+- Do not save one-off completed work, temporary file paths, exact session artifacts, benchmark/task IDs, or command transcripts
+- Save a solution only when it is a reusable technical procedure likely to help with future similar work
+- Prefer concise reproduction steps over a diary of what the agent did
diff --git a/plugins/_memory/prompts/memory.keyword_extraction.sys.md b/plugins/_memory/prompts/memory.keyword_extraction.sys.md
index a0ae25194..33cc8b230 100644
--- a/plugins/_memory/prompts/memory.keyword_extraction.sys.md
+++ b/plugins/_memory/prompts/memory.keyword_extraction.sys.md
@@ -31,23 +31,23 @@ Return ONLY a JSON array of strings containing the extracted keywords/phrases:
## Examples
-**Memory Content**: "Successfully implemented OAuth authentication using JWT tokens for the user login system. The solution handles token refresh and validation properly."
+**Memory Content**: "User prefers Linux shell examples and ./.venv/bin/python paths instead of Windows-only virtualenv commands."
**Output**:
```json
-["OAuth authentication", "JWT tokens", "user login", "token refresh", "authentication implementation"]
+["Linux shell examples", "virtualenv paths", "user preference", "Python commands"]
```
-**Memory Content**: "Fixed the database connection timeout issue by increasing the connection pool size and optimizing slow queries with proper indexing."
+**Memory Content**: "Project uses a configured Dockerized live runtime for smoke checks."
**Output**:
```json
-["database connection", "timeout issue", "connection pool", "query optimization", "indexing"]
+["Dockerized runtime", "live runtime", "smoke checks", "project runtime"]
```
-**Memory Content**: "Learned that Alpine.js x-data components should use camelCase for method names and snake_case for data properties to follow best practices."
+**Memory Content**: "Agent Zero WebUI stores use Alpine.js createStore from /js/AlpineStore.js."
**Output**:
```json
-["Alpine.js", "x-data components", "camelCase methods", "naming conventions"]
+["Agent Zero WebUI", "Alpine.js", "createStore", "AlpineStore.js"]
```
diff --git a/plugins/_memory/prompts/memory.memories_sum.sys.md b/plugins/_memory/prompts/memory.memories_sum.sys.md
index 07ad1abae..82615333b 100644
--- a/plugins/_memory/prompts/memory.memories_sum.sys.md
+++ b/plugins/_memory/prompts/memory.memories_sum.sys.md
@@ -1,10 +1,10 @@
# Assistant's job
1. The assistant receives a HISTORY of conversation between USER and AGENT
-2. Assistant searches for relevant information from the HISTORY worth memorizing
-3. Assistant writes notes about information worth memorizing for further use
+2. Assistant searches for durable information from the HISTORY worth memorizing
+3. Assistant writes notes about stable information worth recalling in future work
# Format
-- The response format is a JSON array of text notes containing facts to memorize
+- The response format is a JSON array of text notes containing durable facts to memorize
- If the history does not contain any useful information, the response will be an empty JSON array.
# Output example
@@ -16,14 +16,18 @@
~~~
# Rules
-- Only memorize complete information that is helpful in the future
+- Only memorize complete information that is likely to remain helpful across future conversations
- Never memorize vague or incomplete information
- Never memorize keywords or titles only
-- Focus only on relevant details and facts like names, IDs, events, opinions etc.
+- Focus on durable user preferences, stable project facts, recurring collaboration constraints, important identities, configured services, and long-lived requirements
+- Do not memorize machine-specific local endpoints, personal absolute paths, container work directories, or ephemeral runtime coordinates as fragment memories; convert to a generic durable fact only if the stable project relationship matters
- Do not include irrelevant details that are of no use in the future
- Do not memorize facts that change like time, date etc.
- Do not add your own details that are not specifically mentioned in the history
- Do not memorize AI's instructions or thoughts
+- Do not memorize what the agent did in this session, commands it ran, files it created, test output, temporary paths, implementation minutiae, or cleanup-only facts
+- Do not memorize a user's one-off request unless it states a durable preference, stable project fact, or recurring constraint
+- If the only notable information is task progress or a completed implementation, return an empty array; reusable procedures belong to successful-solution memory, not fragments
# Merging and cleaning
- The goal is to keep the number of new memories low while making memories more complete and detailed
@@ -36,8 +40,9 @@
# Correct examples of data worth memorizing with (explanation)
> User's name is John Doe (name is important)
-> AsyncRaceError in primary_modules.py was fixed by adding a thread lock on line 123 (important event with details for context)
-> Local SQL database was created, server is running on port 3306 (important event with details for context)
+> User prefers Linux shell commands and relative virtualenv paths over Windows-only examples (stable user preference)
+> Project currently uses a configured live runtime for smoke checks (stable project fact without local coordinates)
+> Runtime-impacting plugin changes must be synced into the configured live environment before testing (recurring project constraint)
# WRONG examples with (explanation of error), never output memories like these
> Dog Information (no useful facts)
@@ -50,6 +55,11 @@
> RAM Status (just a topic without detail)
> User used to prefer X before changing to Y (historical preference is usually not useful; memorize the current preference only)
> Temporary marker ABC123 was used in a memory test (test residue, not useful)
+> Agent created a temporary CLI demo file and ran a shell test (agent action history, not a durable fact)
+> The live UI was reachable at a machine-local endpoint during this session (local runtime detail, not a durable memory)
+> User asked to build a tiny CLI todo app (one-off request, not a recurring preference)
+> The markdown-to-HTML script generated sample.html with 181 bytes (task output, not useful later)
+> AsyncRaceError in primary_modules.py was fixed by adding a thread lock on line 123 (belongs in successful solutions if reusable, not fragments)
# Further WRONG examples
diff --git a/tests/test_memory_quality.py b/tests/test_memory_quality.py
new file mode 100644
index 000000000..d20322ca7
--- /dev/null
+++ b/tests/test_memory_quality.py
@@ -0,0 +1,53 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+if str(PROJECT_ROOT) not in sys.path:
+ sys.path.insert(0, str(PROJECT_ROOT))
+
+from plugins._memory.helpers.memory_quality import (
+ filter_auto_memory_fragments,
+ is_auto_fragment_worth_saving,
+)
+
+
+def test_auto_fragment_quality_keeps_durable_preferences_and_project_facts():
+ assert is_auto_fragment_worth_saving(
+ "User currently prefers concise technical answers with verification."
+ )
+ assert is_auto_fragment_worth_saving(
+ "Project currently uses a configured live runtime for smoke checks."
+ )
+ assert is_auto_fragment_worth_saving(
+ "Runtime-impacting plugin changes must be synced into the configured live environment before testing."
+ )
+
+
+def test_auto_fragment_quality_rejects_action_history_and_transient_artifacts():
+ rejected = [
+ "Agent created a temporary CLI demo file and ran a shell test.",
+ "The user asked to build a tiny CLI todo app in Python.",
+ "Temporary marker ABC123 was used in a memory test.",
+ "The markdown-to-HTML script generated sample.html with 181 bytes.",
+ "Fixed AsyncRaceError in primary_modules.py by adding a thread lock on line 123.",
+ "The live UI was reachable at a machine-local endpoint during this session.",
+ "Project repository path is a personal absolute path on this machine.",
+ ]
+
+ for memory in rejected:
+ assert not is_auto_fragment_worth_saving(memory)
+
+
+def test_auto_fragment_filter_normalizes_and_preserves_kept_order():
+ memories = [
+ "User currently prefers Linux paths in examples.",
+ "Agent created a demo file and reported success.",
+ "Project repository uses a configured source workspace.",
+ ]
+
+ assert filter_auto_memory_fragments(memories) == [
+ "User currently prefers Linux paths in examples.",
+ "Project repository uses a configured source workspace.",
+ ]