35 · Agent Workflow Memory (AWM) — mine reusable recipes from past traces¶

TL;DR. After every solved task, extract a high-level workflow recipe (3-6 generalisable steps). Store recipes in a vector-indexed library. Future tasks retrieve the most-similar recipe and follow it.

Property	Value
Origin	Wang et al., Agent Workflow Memory (2024). arXiv:2409.07429
Storage	Vector-indexed arch.workflows list
Cost	1 retrieve + 1 answer + 1 extract per task = ~3 calls
Sister to	Voyager (nb 29) — skills (code) vs workflows (recipes)

2 · Architecture¶

flowchart LR
    A([task]) --> R[RETRIEVE most-similar workflow]
    R --> AN[ANSWER
prompt prepends recipe if found]
    AN --> EX[EXTRACT new workflow
generalised steps]
    EX --> Z([answer])

    L[(workflow library
vector-indexed)]
    R <-.search.-> L
    EX -.add.-> L

    style R fill:#fff3e0,stroke:#f57c00
    style EX fill:#fce4ec,stroke:#c2185b
    style L fill:#f3e5f5,stroke:#7b1fa2

3 · Theory¶

Voyager (nb 29) stores reusable code (skills). AWM stores reusable strategy (workflows). Skills are concrete; workflows are abstract. For tasks where a small Python function suffices, use Voyager. For tasks that share structure but differ in entities (e.g., "summarise then categorise" applies to news articles AND emails), use AWM.

Demo: 3 sequential tasks of similar structure ("summarise X then categorise it"). Task 1 extracts a workflow; tasks 2 and 3 retrieve and follow it.

4 · Setup¶

In [1]:

from agentic_architectures import get_llm, enable_langsmith, settings
from agentic_architectures.architectures import AgentWorkflowMemory
from agentic_architectures.ui import print_md, print_header
enable_langsmith()
llm = get_llm(provider="nebius", model="meta-llama/Llama-3.3-70B-Instruct", temperature=0.2)
print_header(f"LLM: {llm.model}")

from agentic_architectures import get_llm, enable_langsmith, settings from agentic_architectures.architectures import AgentWorkflowMemory from agentic_architectures.ui import print_md, print_header enable_langsmith() llm = get_llm(provider="nebius", model="meta-llama/Llama-3.3-70B-Instruct", temperature=0.2) print_header(f"LLM: {llm.model}")

LLM: meta-llama/Llama-3.3-70B-Instruct ────────────────────────────────────────────────────────────────────────────

7 · Build the graph¶

In [2]:

from IPython.display import Image, display
arch = AgentWorkflowMemory(llm=llm)
graph = arch.build()
try: display(Image(graph.get_graph().draw_mermaid_png()))
except Exception as e:
    print(f"(PNG unavailable: {e})")
    print(graph.get_graph().draw_mermaid())

from IPython.display import Image, display arch = AgentWorkflowMemory(llm=llm) graph = arch.build() try: display(Image(graph.get_graph().draw_mermaid_png())) except Exception as e: print(f"(PNG unavailable: {e})") print(graph.get_graph().draw_mermaid())

8 · Live run — 3 structurally-similar tasks¶

In [3]:

TASKS = [
    "Summarise this in one sentence and then categorise it (news/opinion/research): 'A new study shows that octopuses may have evolved REM-like sleep states, suggesting that complex dreaming is older than previously thought.'",
    "Summarise this in one sentence and then categorise it (news/opinion/research): 'In my view, the new urban-planning rules will choke small businesses; the city council needs to reconsider before the December deadline.'",
    "Summarise this in one sentence and then categorise it (news/opinion/research): 'After fierce debate, Congress passed the infrastructure bill 218-211 on Tuesday, sending it to the President's desk.'",
]

for tag_idx, q in enumerate(TASKS, 1):
    r = arch.run(q)
    print(f"TASK_{tag_idx}: {q[:80]}...")
    print(f"  USED_RETRIEVED: {r.metadata['used_retrieved_workflow']}")
    print(f"  RETRIEVED_TYPE: {r.metadata['retrieved_workflow_type']!r}")
    print(f"  EXTRACTED_TYPE: {r.metadata['extracted_workflow_type']!r}")
    print(f"  LIBRARY: {r.metadata['library_size_before']} -> {r.metadata['library_size_after']}")
    print(f"  ANSWER: {r.output[:200]}")
    print()
print(f"FINAL_LIBRARY_SIZE: {len(arch.workflows)}")
for w in arch.workflows:
    print(f"  - workflow `{w['task_type']}`:")
    for s in w['steps']: print(f"      • {s}")

TASKS = [ "Summarise this in one sentence and then categorise it (news/opinion/research): 'A new study shows that octopuses may have evolved REM-like sleep states, suggesting that complex dreaming is older than previously thought.'", "Summarise this in one sentence and then categorise it (news/opinion/research): 'In my view, the new urban-planning rules will choke small businesses; the city council needs to reconsider before the December deadline.'", "Summarise this in one sentence and then categorise it (news/opinion/research): 'After fierce debate, Congress passed the infrastructure bill 218-211 on Tuesday, sending it to the President's desk.'", ] for tag_idx, q in enumerate(TASKS, 1): r = arch.run(q) print(f"TASK_{tag_idx}: {q[:80]}...") print(f" USED_RETRIEVED: {r.metadata['used_retrieved_workflow']}") print(f" RETRIEVED_TYPE: {r.metadata['retrieved_workflow_type']!r}") print(f" EXTRACTED_TYPE: {r.metadata['extracted_workflow_type']!r}") print(f" LIBRARY: {r.metadata['library_size_before']} -> {r.metadata['library_size_after']}") print(f" ANSWER: {r.output[:200]}") print() print(f"FINAL_LIBRARY_SIZE: {len(arch.workflows)}") for w in arch.workflows: print(f" - workflow `{w['task_type']}`:") for s in w['steps']: print(f" • {s}")

TASK_1: Summarise this in one sentence and then categorise it (news/opinion/research): '...
  USED_RETRIEVED: False
  RETRIEVED_TYPE: ''
  EXTRACTED_TYPE: 'text_summarization_and_categorization'
  LIBRARY: 0 -> 1
  ANSWER: A new study suggests that octopuses may have evolved REM-like sleep states, indicating that complex dreaming is older than previously thought, and this is categorised as research.

TASK_2: Summarise this in one sentence and then categorise it (news/opinion/research): '...
  USED_RETRIEVED: True
  RETRIEVED_TYPE: 'text_summarization_and_categorization'
  EXTRACTED_TYPE: 'text_classification'
  LIBRARY: 1 -> 2
  ANSWER: The writer argues that new urban-planning rules will harm small businesses and should be reconsidered, which is an opinion.

TASK_3: Summarise this in one sentence and then categorise it (news/opinion/research): '...
  USED_RETRIEVED: True
  RETRIEVED_TYPE: 'text_summarization_and_categorization'
  EXTRACTED_TYPE: 'text_summarization_and_categorization'
  LIBRARY: 2 -> 3
  ANSWER: Congress passed the infrastructure bill 218-211, sending it to the President's desk, which is a news article.

FINAL_LIBRARY_SIZE: 3
  - workflow `text_summarization_and_categorization`:
      • Read and understand the given text
      • Identify the main topic and key points
      • Summarize the text in one sentence
      • Determine the category of the text (news, opinion, or research)
      • Combine the summary and category into a single answer
      • Refine the answer to ensure clarity and coherence
  - workflow `text_classification`:
      • Read the text carefully
      • Identify the main claim or argument
      • Determine the tone and language used
      • Categorise the text as news, opinion, or research
      • Consider the context and potential implications
      • Make a final classification
  - workflow `text_summarization_and_categorization`:
      • Read the given text carefully
      • Identify the main event or topic
      • Determine the category of the text
      • Summarize the text in one sentence
      • Combine the summary and category into a single statement
      • Output the final statement

9 · What we just observed¶

3 sequential summarise-and-categorise tasks. Task 1 extracts a workflow; tasks 2 and 3 should retrieve and follow it.

9.1 · Per-task workflow usage¶

#	Used retrieved?	Retrieved type	Extracted type	Library	Answer
1	❌	—	text_summarization_and_categorization	0→1	A new study suggests that octopuses may have evolved REM-like sleep states, indi
2	✅	text_summarization_and_categorization	text_classification	1→2	The writer argues that new urban-planning rules will harm small businesses and s
3	✅	text_summarization_and_categorization	text_summarization_and_categorization	2→3	Congress passed the infrastructure bill 218-211, sending it to the President's d

9.2 · The workflow library that built up¶

text_summarization_and_categorization

• Read and understand the given text
• Identify the main topic and key points
• Summarize the text in one sentence
• Determine the category of the text (news, opinion, or research)
• Combine the summary and category into a single answer
• Refine the answer to ensure clarity and coherence

text_classification

• Read the text carefully
• Identify the main claim or argument
• Determine the tone and language used
• Categorise the text as news, opinion, or research
• Consider the context and potential implications
• Make a final classification

text_summarization_and_categorization

• Read the given text carefully
• Identify the main event or topic
• Determine the category of the text
• Summarize the text in one sentence
• Combine the summary and category into a single statement
• Output the final statement

9.3 · Patterns surfaced¶

• ✅ Task 2 reused task 1's workflow (text_summarization_and_categorization). Memory pipeline alive.

• ✅ Task 3 reused a workflow (text_summarization_and_categorization).

• Final library size: 3 workflow(s) extracted across 3 task(s).

9.4 · The takeaway¶

AWM's value is in the Used retrieved? column of § 9.1. The first task always extracts (no library yet); subsequent structurally-similar tasks should retrieve and use the prior workflow. If the column reads ✅, ❌, ❌ across 3 similar tasks, vector retrieval is broken; if ✅, ✅, ✅ it's working — the library is amortising strategy cost across calls.

11 · Failure modes & extensions¶

| Failure | Mitigation | |---|---| | Workflow too specific | Steps reference task-1 entities; tasks 2-3 don't fit | Strict prompt rule in extraction: "no specific entity names" | | Workflow library bloat | N tasks → N near-duplicate workflows | Cosine-dedup on workflow descriptions | | Retrieved workflow misfires | Wrong workflow retrieved; answer follows bad recipe | Add post-answer verification step |

Extensions: (1) Combine with Voyager skills (workflow specifies which skills to call), (2) hierarchical workflows (workflow contains sub-workflows), (3) workflow scoring by usefulness across tasks.

Reference: Wang et al., AWM. 2024. arXiv:2409.07429