REVIEW R. Menzel Memory dynamics in the honeybee Accepted: 12 May 1999 Abstract Reward learning in honeybees initiates a se- quence of events which leads to long-lasting memory passing through multiple phases of transient memories. The study of memory dynamics is performed at the be- havioral (both natural foraging behavior and appetitive conditioning), neural circuit and molecular levels. The results of these combined e��orts lead to a model which assumes five kinds of sequential memories, each characterized by a set of behavioral and mechanistic properties. It is argued that these properties, although reflecting general characteristics of step-wise memory formation, are adapted to the species-specific adapta- tions in natural behavior, here to foraging at scattered and unreliable food sources. Key words Learning AE Memory phases AE Foraging AE Neural circuits AE Signal cascades Abbreviations AL antennal lobe AE CS conditioned stimulus AE eSTM early short-term memory AE ITI intertrial interval AE lACT lateral antenno-calycal tract AE lSTM late short-term memory AE LTM long-term memory AE mACT median antenno-calycal tract AE mb mushroom bodies AE mlACT medio-lateral antenno- calycal tract AE MTM mid-term memory AE OA octopamine AE PER proboscis extension response AE PKA cAMP-dependent protein kinase A AE PKC Ca2+/ phospholipid-dependent protein kinase C AE US unconditioned stimulus Introduction Learning leads to memory, but memories are not made instantaneously by learning rather, they develop over time and change their properties. When Hermann Ebbinghaus began the scientific study of memory in humans in l885, he discovered that memory formation is a time-consuming process and depends on the interval between sequential learning trials. Later, James (1890) developed the concept of primary and secondary mem- ory, referring to the limited time span and information capacity of primary memory, and the seemingly unlim- ited duration and content of secondary memory. Around the turn of the century, psychologists had es- tablished a framework of thinking about sequential memory stages, which was captured by the preservation- consolidation hypothesis of Muller and Pilzecker (1900): neural processes underlying newly-formed memories initially persevere in a labile form and then, over time, become consolidated into lasting neural traces. Considerable e��ort has been put into identifying the neural and cellular substrates of these di��erent forms of memory, and on the whole, the hypothesis that memory storage processes are time dependent has been most fruitful (Squire 1987 Kandel and Squire 1992). But why should memory take hours, days or weeks for final adjustment of the circuit? Is the neural ma- chinery so slow? This is rather unlikely, given the speed with which neural excitation can lead to gene activation, protein synthesis, protein translocation along axons and dendrites, and finally, restructuring of synaptic contacts (Frey and Morris 1997), and given the fact that certain important learning tasks lead to stable memory very quickly. I want to argue here that the dynamics of memory traces need to be seen in the context of natural behavior under which memory formation takes place. Perseveration and consolidation of memories are adapted to the demands and constraints of the specific requirements to which the animal is exposed in its nat- ural environment. Animals behave at any certain time with reference to information gathered over long periods of time, and at any particular moment information is evaluated according to genetically-controlled internal conditions (thus the phylogenetic history of the species) and parameters of the experience-gathering process such J Comp Physiol A (1999) 185: 323���340 �� Springer-Verlag 1999 This paper is dedicated to Martin Lindauer on his 80th birthday R. Menzel Institut fur Neurobiologie, Freie Universitat Berlin, Konigin-Luise-Strasse 28/30, D-14195 Berlin, Germany

as reliability, context-dependence and the meaning of the new information to the animal. These aspects are a function of time. New memories have to be incorporated into existing ones on the basis of their relevance. It is fair to assume that restructuring circuits is costly to the nervous system, thus active processes might keep the neural trace of new memories from being fixed in structure as long as they are not finally evaluated by additional information resulting from ongoing behavior (Yin et al. 1995a, b). Suppres- sion of memory formation may thus be an equally im- portant process in the timing of memory phases, and the balance between activation and suppression of memory formation processes may carry the secret for under- standing the dynamics of memory formation. In an attempt to unravel these processes it might help to analyze the dynamics of memory from many vantage points and to incorporate the biological context in which they have been developed. We have chosen foraging behavior of honeybees as a study case. Here I shall demonstrate that (1) bees do indeed have memory phases, (2) that these memory phases can be correlated with stages in the foraging cycle, and (3) that the memories have di��erent properties with respect to their localization in the brain, the control of behavior, and the cellular reaction cascades underlying these memory phases. The dynamics of the memory phases will be analyzed at three levels ��� behavioral, neural and cellular. Behavior Foraging behavior Bees learn flowers��� local cues (color, odor, shape, loca- tion, handling) quickly and e��ectively according to the reward (nectar, pollen) they provide. Flowers as food sources are unreliable and widely scattered. Each flower provides only a minute amount of reward. The profit- ability of a plant species is defined by the amount of reward obtained relative to the e��orts invested to collect it (Nunez and Giurfa 1996), and it is profitability which guides bees in their foraging behavior (Greggers and Menzel 1993). A foraging bout is structured in time (Fig. 1). Because flowers appear in patches, intrapatch choices follow each other quickly and are more likely to hit on the same kind of flower. Interpatch choices are more spread out in time, and are likely to expose bees to other flowers, forcing them to make decisions between ������same������ and ������di��erent������ flowers. Interbout intervals can vary between several minutes and months, e.g., when an overwintered bee flies out for the first time in the spring. Memories steering the choice behavior may di��er ac- cordingly. An early short-term memory (eSTM) (or ������working memory������) keeps an active memory over short periods of time, late short-term memory (lSTM) may include short-term retrieval processes facilitating the decision between similar and di��erent, and bridges the gap between short-term memory and longer-lasting forms of memory. These are likely to be structured in di��erent phases because they lead to other forms of memory which control behavior over a wide range of time (hours to months). The longer-lasting forms also refer to memory which needs to be retrieved from a re- mote store to become e��ective, and any new learning may not have direct access to long-term storage, but rather has to be stored in lSTM for processing which then leads to an addition of these mid- and long-term memories. This temporal structure of the foraging cycle may provide a framework for understanding the se- quence and properties of memory phases in bees. The temporal sequence of choices in the short-term range during foraging behavior under natural conditions does indeed indicate the sequence of two di��erent forms of memory (Fig. 2 Chittka et al. 1997). Individual ani- mals of four di��erent species of bumble bees were re- corded for their sequential choice behavior between five di��erent plant species flowering in an area of 20 m �� 8 m. The frequency of intervals between choices of the same plant species (stay flights) is di��erent from that of choices of di��erent plant species (shift flights). Fig. 1 Foraging cycle in honeybees. Flowers usually appear in patches of similar flowers. A bee departing from the hive will arrive at such a patch, and make intra-patch choices at short intervals. Inter- patch choices follow each other at longer intervals and require a decision between similar and di��erent flowers. Intervals between bouts range among many minutes, hours and days (Menzel 1985). The lower figure depicts a working hypothesis of di��erent memories as defined by the sequences of events during a natural foraging cycle. eSTM early short-term memory, lSTM late short-term memory, MTM mid- term memory, LTM late short-term memory 324

Same flowers are chosen more frequently at short in- tervals, di��erent flowers at longer intervals. This dis- tinction between stay and shift flights applies irrespective of the density and distribution of the flowers. This be- comes particularly evident for flowers of two plant species (Lotus corniculatuus, Lathyrus pratensis) which have the same colors for bees and therefore cannot be discriminated. These flowers are chosen as if they were the same irrespective of their distribution. The choices between them follow the distribution for stay flights. These results indicate that individual bumble bees which forage on several plant species simultaneously neither distribute their choices randomly nor do they follow a simple performance rule such as ������fly straight ahead if little reward is experienced and fly a curve if a high reward is experienced������ (Pyke 1984). Instead, their choice performance is guided by their short-term memories (eSTM and lSTM). It is, however, unknown why bees sometimes make decisions at shorter and sometimes at longer intervals. The amount of reward experienced at the flower may play an important role, and it has indeed been shown that bees develop reward-specific memories for feeders di��ering in the amount of reward (Greggers and Menzel 1993 Greggers and Mauelshagen 1997). Thus, it might be that bees develop ������expectancies������ for the reward quality of food sources and guide their choice behavior according to the contents of their sequential memories. This question can be studied under semi-natural ex- perimental conditions which allow full control over the amount of reward and the sequential behavior. Such a situation is described in Fig. 3. A bee forages in a patch of four electronic feeders which provide sucrose at a low, constant flow rate. Under such conditions bees visit each feeder with equal frequency, but not randomly (Ma- uelshagen and Greggers 1993). Greggers and Mauelsh- agen (1995) found that the lick time during each visit is a measure of what the animals recall about the reward from the feeder, because the lick time during the last visit to the same feeder correlates highly with the actual lick time (Fig. 3Bb same feeder). No such correlation is found when di��erent feeders are compared (Fig. 3Bb di��erent feeder). Thus, correlation between actual and last lick time can be taken as a measure of reward memory. If sequential visits (Fig. 3A next choice) over short intervals (1 min) are considered, memory decays quickly and no di��erence is found between stay and shift flights (Fig. 3Ab), indicating that an unspecific memory contributes to the behavior during the time immediately following learning. If longer intervals are considered during which the bee performs one visit to another feeder (Fig. 3B next choice but one), a di��erent time function is found for the second subsequent choice: memory for the same feeder improves over time, and becomes increasingly specific because transfers to dif- ferent feeders no longer occur (Fig. 3Bc). We conclude from these results that eSTM can be separated from lSTM. eSTM appears as less specific than lSTM, and lSTM develops over time in the minute range, leading to more specific memory content. So far we have considered early memories during appetitive learning. Long-term memories are known to guide bee foraging (Frisch 1967) and can be separated from lSTM (Menzel 1990). Lindauer (1963) reported that bees remember the color of a feeding place over several months. Flower fidelity (frequently documented in honeybees) is only possible if they control their for- aging behavior by long-lasting memories. Dual choice color discrimination A single learning trial in a dual-choice color learning task initiates time-dependent processes which resemble Fig. 2A, B Temporal dynamics of flower choices in 4 bumblebee species under natural conditions. Five di��erent plant species flowered in the study area (20 m �� 8 m). A Schematic distribution of the five species of flowers, and their relative frequency in numbers, respec- tively. The arrows indicated the two species whose flowers cannot be discriminated by the bumblebees on the basis of their floral color. B Temporal dynamics of stay and shift flights in functions normalized to their respective maxima. Stay flights (choice of the same plant species) follow each other at shorter intervals than shift flights (choice of a di��erent plant species). The distribution and frequency of the flowers does not explain the time-course of stay and shift flights. (adapted from Chittka et al. (1997), averaged, normalized and replotted) 325