KurisutinaResearch date: 2026-09-22.
Citation and status. Sean M. Polyn, Kenneth A. Norman, and Michael J. Kahana. A context maintenance and retrieval model of organizational processes in free recall. Peer-reviewed, Psychological Review 116(1), 129–156, January 2009. DOI; bibliographic record; public author manuscript.
Acquisition and reading. Saved PDF, readable text, and provenance and hashes. Read all 32 manuscript pages: main text, model equations, results, discussion, figure legends, references, and Appendices A–C covering experimental methods, semantic calibration, and fitting. This is the author's in-press version dated 2008-10-29, not a verified identical copy of the final typeset article. No external code, raw data, or subsequent replication was inspected. No open redistribution license was identified. Download transport details are recorded in provenance.
Source findings and limits (factual paraphrase). Forty-five adults studied 24-word lists using size or animacy judgments, then recalled freely. Context Maintenance and Retrieval (CMR) models each recall as a noisy competition, followed by reinstatement of associated context that biases subsequent retrieval. It combines semantic associations, temporal context, source context, and disruption at task changes. Thirteen parameters were fitted to 93 aggregate summaries from the new experiment; separate fits covered two earlier datasets. The full model outperformed two restricted variants, while still deviating significantly from the data. Source-switch latency was not fitted: its predicted direction matched observations, but the full model's cost was 432 milliseconds versus 1,302 observed. This check used the same experimental dataset, not an independent replication. Individuals were not fitted; semantic variation was approximated using population associations and a noise correction. The work establishes a useful behavioral account, not unique recovery of an individual's memory structure or identification of a neural mechanism.
Operational interpretation — our analysis of the model. The essential loop is:
[ \text{context};c_t \longrightarrow \text{item support};M^{CF}c_t \longrightarrow \text{recalled item};f_t \longrightarrow \text{updated context};c_{t+1}. ]
Equation 2 combines retained context with context supplied by the item, (c_{t+1}=\rho c_t+\beta c^{IN}); normalization and separate context components matter. Thus an observed transition is not simply an edge being read out. It is the outcome of an interaction among stored associations, current context, competitors, and a stopping/decision process. The scientific question is which parts of that interaction are needed to forecast behavior beyond the observations used to construct them. Here the candidate items are supplied by the studied list; discovering an unknown autobiographical candidate set remains a separate problem. Instructed list recall also requires separate validation before generalizing to involuntary autobiographical recollection.
Comparison with a static graph — our analysis. Define the baseline carefully. A fixed transition graph whose next-step probabilities depend only on the latest item and which items remain available has no representation of the order of earlier recalls. A maintained context state can preserve such information. But a fixed graph with evolving activation, recency traces, or history-dependent transition weights is already a dynamical system. Beating the restricted baseline would not establish that graphs are inadequate.
A simple algebraic example illustrates the distinction. Consider our deliberately simplified update (c_{t+1}=a c_t+bVf_t), with fixed scalar coefficients and item support (u_t=Wc_t). Defining the graph (G=WV) yields (u_{t+1}=a u_t+bGf_t). A fixed association graph plus a changing support vector can therefore realize this simplified readout exactly. This is not a reduction of full CMR, whose normalization, multiple context components, and competitive decision process must also be represented. It demonstrates why storage format alone is an insufficient scientific contrast.
What recall sequences identify — our reasoning. Rich sequences constrain conditional behavior more strongly than an unordered set of recalled items. They still do not, by themselves, separate strong associations from a favorable starting context, weak competition, an unobserved rehearsal, or a deliberate search strategy. Similar average clustering can arise through different combinations. An individual's fitted parameter should therefore be treated as a model-dependent estimate, accompanied by uncertainty and parameter-recovery checks, rather than a discovered psychological trait. Population semantic norms also need a matched individual-information baseline before attributing a gain to better dynamics.
Falsifiable experiment — proposal. First study controlled material with independently varied semantic similarity, temporal proximity, and source task. After encoding, randomize two cue sequences containing exactly the same seed items in different orders, with the same final seed. Use fresh matched lists to avoid testing a supposedly untouched memory twice. This fixes the remaining candidate set and last cue while manipulating the preceding retrieval context. Exposure, cue timing, and reporting instructions should be matched; randomize the manipulation after encoding so initial learning cannot explain a condition difference. Cued reactivation is itself an intervention, so the result concerns that intervention's effect on subsequent free recall, not unobserved spontaneous thought.
Freeze models before test lists. Compare a population baseline, a personalized fixed transition graph, that same graph with explicit decaying history, and a fitted context model, all supplied the same observations. Score probabilities of next item, stopping, and response-onset latency, plus complete-sequence summaries on separate held-out episodes. Reporting onset does not directly measure covert retrieval timing. Conditional next-response prediction and autonomous sequence generation should both be evaluated: matching one does not guarantee the other. A context model gains substantive support if it forecasts the direction and magnitude of history effects across new materials or task schedules. Equal performance from a simpler history model leaves its additional latent machinery unsupported.
Implication for extraction — proposal. Preserve cue order, retrieval timing, task instructions, intervening activities, and omissions along with recalled content. A cue that changes the next several recollections may be useful for extraction while also changing the process being measured. Evaluate episode coverage separately from fidelity to the person's unaided recall distribution; neither is a substitute for the other.