Most "AI persona" tools have no memory. Each interview session starts with the model rehydrating a character description and improvising fresh answers. Two sessions with the same character produce subtly different people because the underlying context resets. The character doesn't remember what it said last week, can't reference its own past statements, and can't carry beliefs forward when those beliefs were formed during a prior conversation. The result is a fluent character that fails the most basic test of being a respondent: actually being the same person across multiple conversations.
This essay walks through how Candor's participant memory architecture works. Six memory types, each grounded in a documented cognitive science framework, together making a synthetic participant behave as the same person across every interview session within a study. If you want the opinionated version of why this matters, read why synthetic research needs evidence grounding. This piece is the technical companion focused specifically on the memory layer. For the broader Candor walk-through, see how Candor works.
What participant memory means in practice
A synthetic respondent without memory is a character. A synthetic respondent with memory is a respondent. The distinction matters because real research depends on continuity. When a UX researcher schedules a follow-up interview with a real customer two weeks later, the customer remembers what they said last time, can refer back to specific decisions they described, and updates their views in light of new information without forgetting where they started. Without that, longitudinal research isn't possible.
Candor's participant memory architecture exists to make that same property available for synthetic respondents within a study. A participant interviewed today can be re-interviewed next week with full context preserved. Their stated beliefs persist. Their described behaviors persist. Their past decisions are documented and queryable. Their language style is consistent. The conversation history is referenceable. Memory makes the participant a research subject rather than a fresh character generated per session.
Six memory types populate when the participant is created, and most of them update as interviews accumulate. The six are grounded in distinct cognitive science research traditions, which is the right shape because human memory itself is not one undifferentiated system. The next sections walk through each.
The six memory types
Identity memory. Who the participant is. Name, demographic background, professional role, life stage, region, relevant biographical context. Identity memory is their "control structure" in Conway's Self-Memory System terms: it organizes all other memory access and changes slowly over time. A participant's identity stays stable across the study; they don't become someone different between interview sessions. Research grounding: Conway's working self + conceptual self, plus McAdams' Level 3 (narrative identity core).
Behavioral memory. What the participant actually does. Routines, purchase patterns, work habits, app usage, communication preferences, decision-making sequences. Behavioral memory captures their documented patterns of action, not abstract attitudes. Research grounding: Tulving's semantic + procedural memory; Conway's general events level; Schank/Abelson scripts (the routine patterns humans use to navigate familiar situations without re-reasoning each time).
Belief memory. What the participant believes to be true. Attitudes toward categories, products, brands, problems, solutions. Values and assumptions about the world. Belief memory is updatable through stance changes, but the update rule is conservative: prior beliefs are preserved alongside new ones, mirroring how humans retain awareness of "what I used to think" even after updating. Research grounding: Conway's conceptual self (attitudes, values); McAdams' Level 3 (thematic threads); Schank/Abelson plans (evaluative frameworks).
Language memory. How the participant talks. Vocabulary, idioms, register, formality level, professional jargon, communication style. Language memory ensures they sound like the same person across sessions, with consistent ways of describing problems and reacting to questions. Research grounding: Tulving's procedural memory (language production runs on procedural systems); Conway's conceptual self (self-presentation through language); McAdams' Level 1 (dispositional expression).
Decision memory. How the participant has decided things. Each significant decision they describe in an interview is logged with the criteria used, the trade-offs considered, the alternatives evaluated, and the outcome. Decision memory is append-only: each decision is a unique episode and cannot be retroactively rewritten. Research grounding: Tulving's episodic memory; Conway's event-specific knowledge; McAdams' Level 2 (contextualized choices).
Conversation memory. What the participant has said in past interview sessions. Conversation memory starts empty when they are created and accumulates with each interview. Like decision memory, it's append-only. Every prior session is a distinct episode in their autobiographical timeline, and the current session can reference any of them. Research grounding: Tulving's episodic memory again, but at the dialog level; McAdams' Level 3 (narrative episodes); Schank/Abelson script deviations and novel encounters.
These six map to documented memory systems in cognitive science, not to convenience categories. The mapping isn't arbitrary; it reflects how human memory actually divides into stable identity, semantic knowledge of self and world, procedural skill, and episodic memory of specific events. Treating synthetic memory as a single bucket would lose the architectural distinctions that make the system behave like a respondent rather than a character.
How memory persists across sessions
A participant generated for a study lives across every interview session in that study. Generated once. Interviewed many times. The memory layer is what makes that work.
When a new interview session begins, the participant's memory base is loaded into context: identity, behaviors, beliefs, language style, decisions made in prior sessions, and conversation history from prior sessions. The interviewer-side prompt frames them as the same person being re-engaged, not a fresh character being instantiated. Their responses draw on the loaded memory base; references to "last time we spoke" or "as I mentioned before" are grounded in actual prior session content, not improvised.
After the session completes, the relevant memory types update. New decisions get appended to decision memory. Stance changes (genuine belief updates triggered by the interview) get logged to belief memory while preserving the prior belief. The session itself becomes a new entry in conversation memory. Behavioral memory updates if the participant described a new routine. Language memory refines marginally based on the session's communication patterns. Identity memory stays stable; they don't become a different person between sessions.
The update rules are deliberately conservative. Identity memory rarely changes. Belief memory updates only on documented stance changes, never on idle drift. Decision memory and conversation memory are append-only; you can never rewrite a past decision or erase a past session. The conservatism is grounded in Conway's self-coherence finding: humans resist belief change to maintain identity continuity, and rapid identity change would destabilize the memory system. Candor's update rules mirror that property.
The result is a participant who behaves as the same person across the entire study, with documented history that's referenceable in any later session and a belief structure that can update when warranted without losing its prior state.
How memory interacts with personality and biases
Memory is what the participant reasons about. Personality and cognitive biases are how they reason. The two layers are orthogonal but work together at runtime.
A participant with high Conscientiousness and high loss-aversion will reason about a decision in their decision memory differently from one with low Conscientiousness and low loss-aversion, even when both have the same prior decision in memory. The OCEAN profile and bias profile shape the reasoning approach. Memory provides the substrate the reasoning operates on. Without memory, the personality has nothing specific to reason about. Without personality and biases, everyone reasons about the same memory base identically, which would collapse the differences between groups.
For the methodology layer on personality, see the OCEAN model in synthetic participants. The two essays describe complementary layers of the same model: personality and biases shape the reasoning style; memory provides the content to reason about.
What memory does not capture
A memory architecture grounded in cognitive science still has documented limits. Three categories of thing memory does not capture, and acknowledging them is part of the methodology.
Actual lived experience. Candor participants are evidence-grounded constructs derived from published research, customer-uploaded documents, and validated distributions. Their behavioral memory reflects documented patterns of the audience they were generated to represent; it isn't a record of any specific real person's actual life. Memory makes a participant consistent across sessions; it doesn't make them a real human. This limit is the same one that applies to synthetic research generally and is addressed honestly in the methodology rather than papered over.
Observed in-product behavior. A participant's behavioral memory captures documented patterns of the audience, not real-time observations of a specific human navigating a real product. Real product analytics and tracking studies capture that data; a participant doesn't substitute for it. For research questions where ground-truth in-product behavior is the answer, panel research or real-user testing remains required.
Genuine emotional valence beyond what's modeled. Memory captures what the participant has said and decided. Personality and bias profiles capture how they reason. Neither layer captures genuine emotional experience the way a real respondent's account does. The participant reasons about emotional states reported in the source evidence; they don't experience them. This is a documented limit of synthetic research as a methodology and applies to every layer of the model, not just memory.
Cross-study memory. Participants are scoped to a single study. One generated for Study A doesn't carry memory into Study B. This is by design: each study has its own audience, learning goals, and evidence pool, so participants are regenerated from the study-specific evidence base rather than reused across contexts. The cross-session continuity is within a single study, not across the operation as a whole. For research that needs to track one participant's evolution across multiple research contexts, that's a future-direction conversation rather than a shipping capability.
Operational implications for research practice
The memory architecture changes what's possible in a research workflow. Three concrete shifts worth naming.
Longitudinal research within a study becomes practical. A researcher running problem discovery this week can return to the same participants next week for concept testing, with full memory of the prior conversation. They reference what they said before, update their views based on what was raised, and carry context forward. Traditional synthetic research that resets between sessions can't do this. Real-customer research can, but at the cost of weeks of recruitment per round.
Probing depth increases over sessions. Because the participant remembers prior context, the researcher doesn't have to re-establish baseline information each time. The second and third sessions can go deeper into specifics the first session surfaced. The cumulative depth per participant is higher than what a single one-shot interview can produce.
Belief drift becomes trackable. Stance changes are logged to belief memory when they occur. Researchers can see which beliefs the participant updated, what triggered the update, and what the prior belief was. The drift is a research signal in its own right, especially in problem validation, where the question is whether a stated problem holds up under probing.
The architectural feature that distinguishes a character you talk to once from a synthetic respondent you can develop a research relationship with is memory. It's the closest thing in the synthetic research stack to the actual property of real human respondents that makes longitudinal research possible.
For the pieces that pair with this one, see how evidence grounding works for the upstream pipeline that builds the evidence pool memory draws on, how group clustering works for the step that turns evidence into the group templates participants are sampled from, and the OCEAN essay linked above for the personality layer memory works with.