Survey Memory System

Memory controls which previous question-answer pairs an agent sees when answering each question. This affects how context is provided in the LLM prompt.

Memory Modes

Mode	Description	Use Case
No Memory	Agent sees no prior answers	Independent questions
Full Memory	Agent sees ALL prior answers	Conversational surveys
Lagged Memory	Agent sees last N answers	Recent context only
Targeted Memory	Specific questions see specific priors	Precise control

Full Memory Mode

Every question sees all prior answers:

python

from edsl import Survey

survey = Survey([q1, q2, q3, q4, q5])

# Enable full memory (returns new Survey)
survey = survey.set_full_memory_mode()

# Now:
# - q2 sees q1's answer
# - q3 sees q1 and q2's answers
# - q4 sees q1, q2, and q3's answers
# - etc.

Lagged Memory Mode

Each question sees only the last N answers:

python

survey = Survey([q1, q2, q3, q4, q5])

# Each question sees the last 2 Q&A pairs
survey = survey.set_lagged_memory(lags=2)

# Now:
# - q3 sees q1 and q2's answers
# - q4 sees q2 and q3's answers
# - q5 sees q3 and q4's answers

Targeted Memory

Specific questions see specific prior answers:

python

survey = Survey([q1, q2, q3, q4, summary])

# q3 remembers q1's answer
survey = survey.add_targeted_memory("q3", "q1")

# summary remembers multiple specific questions
survey = survey.add_memory_collection("summary", ["q1", "q2", "q3"])

# Chain multiple targeted memories
survey = (survey
    .add_targeted_memory("q3", "q1")
    .add_targeted_memory("q4", "q2")
    .add_memory_collection("summary", ["q1", "q2", "q3", "q4"]))

Memory Collection

One question sees multiple specific prior answers:

python

# Create a summary question that sees specific prior answers
summary_q = QuestionFreeText(
    question_name="summary",
    question_text="Based on your previous answers, what patterns do you notice?"
)

survey = Survey([q1, q2, q3, q4, summary_q])

# summary sees answers from q1, q2, q3, q4
survey = survey.add_memory_collection("summary", ["q1", "q2", "q3", "q4"])

Combining Memory Modes

You can combine different memory approaches:

python

survey = (Survey([intro, q1, q2, q3, q4, summary])
    # Start with lagged memory as base
    .set_lagged_memory(lags=1)
    # Add specific memory for summary
    .add_memory_collection("summary", ["q1", "q2", "q3", "q4"]))

Memory and Dependencies

Memory creates dependencies in the survey's DAG (Directed Acyclic Graph):

•If question B remembers question A, then B depends on A
•This affects parallel execution and validation
•Questions cannot remember future questions (forward references)

python

# This works - q3 remembers q1
survey = survey.add_targeted_memory("q3", "q1")

# This would fail - can't remember future questions
# survey = survey.add_targeted_memory("q1", "q3")  # Error!

View Memory Configuration

python

# Access the memory plan
memory_plan = survey.memory_plan

# See memory for each question
for question_name, memory in memory_plan.data.items():
    print(f"{question_name} remembers: {memory.data}")

How Memory Affects Prompts

When an agent answers a question with memory, the prompt includes:

code

Before this question, you already answered:

Question: What is your name?
Answer: Alice

Question: What is your occupation?
Answer: Software engineer

Now answer this question:
...

Memory Modes Comparison

python

survey = Survey([q1, q2, q3, q4])

# No memory (default) - each question is independent
# q4 sees: nothing

# Full memory
survey_full = survey.set_full_memory_mode()
# q4 sees: q1, q2, q3

# Lagged memory (lags=2)
survey_lagged = survey.set_lagged_memory(lags=2)
# q4 sees: q2, q3

# Targeted memory
survey_targeted = survey.add_targeted_memory("q4", "q1")
# q4 sees: q1 only

Quick Reference

Task	Method
Full memory	`survey.set_full_memory_mode()`
Lagged memory	`survey.set_lagged_memory(lags=N)`
Single targeted	`survey.add_targeted_memory(focal, prior)`
Multiple targeted	`survey.add_memory_collection(focal, [priors])`
View memory plan	`survey.memory_plan`